Kernel Probes (Kprobes)

Author

Jim Keniston <jkenisto@us.ibm.com>

Author

Prasanna S Panchamukhi <prasanna.panchamukhi@gmail.com>

Author

Masami Hiramatsu <mhiramat@redhat.com>

Concepts: Kprobes and Return Probes

Kprobes enables you to dynamically break into any kernel routine and collect debugging and performance information non-disruptively. You can trap at almost any kernel code address 1, specifying a handler routine to be invoked when the breakpoint is hit.

1

some parts of the kernel code can not be trapped, see Blacklist)

There are currently two types of probes: kprobes, and kretprobes (also called return probes). A kprobe can be inserted on virtually any instruction in the kernel. A return probe fires when a specified function returns.

In the typical case, Kprobes-based instrumentation is packaged as a kernel module. The module’s init function installs (“registers”) one or more probes, and the exit function unregisters them. A registration function such as register_kprobe() specifies where the probe is to be inserted and what handler is to be called when the probe is hit.

There are also register_/unregister_*probes() functions for batch registration/unregistration of a group of *probes. These functions can speed up unregistration process when you have to unregister a lot of probes at once.

The next four subsections explain how the different types of probes work and how jump optimization works. They explain certain things that you’ll need to know in order to make the best use of Kprobes – e.g., the difference between a pre_handler and a post_handler, and how to use the maxactive and nmissed fields of a kretprobe. But if you’re in a hurry to start using Kprobes, you can skip ahead to Architectures Supported.

How Does a Kprobe Work?

When a kprobe is registered, Kprobes makes a copy of the probed instruction and replaces the first byte(s) of the probed instruction with a breakpoint instruction (e.g., int3 on i386 and x86_64).

When a CPU hits the breakpoint instruction, a trap occurs, the CPU’s registers are saved, and control passes to Kprobes via the notifier_call_chain mechanism. Kprobes executes the “pre_handler” associated with the kprobe, passing the handler the addresses of the kprobe struct and the saved registers.

Next, Kprobes single-steps its copy of the probed instruction. (It would be simpler to single-step the actual instruction in place, but then Kprobes would have to temporarily remove the breakpoint instruction. This would open a small time window when another CPU could sail right past the probepoint.)

After the instruction is single-stepped, Kprobes executes the “post_handler,” if any, that is associated with the kprobe. Execution then continues with the instruction following the probepoint.

Changing Execution Path

Since kprobes can probe into a running kernel code, it can change the register set, including instruction pointer. This operation requires maximum care, such as keeping the stack frame, recovering the execution path etc. Since it operates on a running kernel and needs deep knowledge of computer architecture and concurrent computing, you can easily shoot your foot.

If you change the instruction pointer (and set up other related registers) in pre_handler, you must return !0 so that kprobes stops single stepping and just returns to the given address. This also means post_handler should not be called anymore.

Note that this operation may be harder on some architectures which use TOC (Table of Contents) for function call, since you have to setup a new TOC for your function in your module, and recover the old one after returning from it.

Return Probes

How Does a Return Probe Work?

When you call register_kretprobe(), Kprobes establishes a kprobe at the entry to the function. When the probed function is called and this probe is hit, Kprobes saves a copy of the return address, and replaces the return address with the address of a “trampoline.” The trampoline is an arbitrary piece of code – typically just a nop instruction. At boot time, Kprobes registers a kprobe at the trampoline.

When the probed function executes its return instruction, control passes to the trampoline and that probe is hit. Kprobes’ trampoline handler calls the user-specified return handler associated with the kretprobe, then sets the saved instruction pointer to the saved return address, and that’s where execution resumes upon return from the trap.

While the probed function is executing, its return address is stored in an object of type kretprobe_instance. Before calling register_kretprobe(), the user sets the maxactive field of the kretprobe struct to specify how many instances of the specified function can be probed simultaneously. register_kretprobe() pre-allocates the indicated number of kretprobe_instance objects.

For example, if the function is non-recursive and is called with a spinlock held, maxactive = 1 should be enough. If the function is non-recursive and can never relinquish the CPU (e.g., via a semaphore or preemption), NR_CPUS should be enough. If maxactive <= 0, it is set to a default value. If CONFIG_PREEMPT is enabled, the default is max(10, 2*NR_CPUS). Otherwise, the default is NR_CPUS.

It’s not a disaster if you set maxactive too low; you’ll just miss some probes. In the kretprobe struct, the nmissed field is set to zero when the return probe is registered, and is incremented every time the probed function is entered but there is no kretprobe_instance object available for establishing the return probe.

Kretprobe entry-handler

Kretprobes also provides an optional user-specified handler which runs on function entry. This handler is specified by setting the entry_handler field of the kretprobe struct. Whenever the kprobe placed by kretprobe at the function entry is hit, the user-defined entry_handler, if any, is invoked. If the entry_handler returns 0 (success) then a corresponding return handler is guaranteed to be called upon function return. If the entry_handler returns a non-zero error then Kprobes leaves the return address as is, and the kretprobe has no further effect for that particular function instance.

Multiple entry and return handler invocations are matched using the unique kretprobe_instance object associated with them. Additionally, a user may also specify per return-instance private data to be part of each kretprobe_instance object. This is especially useful when sharing private data between corresponding user entry and return handlers. The size of each private data object can be specified at kretprobe registration time by setting the data_size field of the kretprobe struct. This data can be accessed through the data field of each kretprobe_instance object.

In case probed function is entered but there is no kretprobe_instance object available, then in addition to incrementing the nmissed count, the user entry_handler invocation is also skipped.

How Does Jump Optimization Work?

If your kernel is built with CONFIG_OPTPROBES=y (currently this flag is automatically set ‘y’ on x86/x86-64, non-preemptive kernel) and the “debug.kprobes_optimization” kernel parameter is set to 1 (see sysctl(8)), Kprobes tries to reduce probe-hit overhead by using a jump instruction instead of a breakpoint instruction at each probepoint.

Init a Kprobe

When a probe is registered, before attempting this optimization, Kprobes inserts an ordinary, breakpoint-based kprobe at the specified address. So, even if it’s not possible to optimize this particular probepoint, there’ll be a probe there.

Safety Check

Before optimizing a probe, Kprobes performs the following safety checks:

• Kprobes verifies that the region that will be replaced by the jump instruction (the “optimized region”) lies entirely within one function. (A jump instruction is multiple bytes, and so may overlay multiple instructions.)

• Kprobes analyzes the entire function and verifies that there is no jump into the optimized region. Specifically:

  • the function contains no indirect jump;

  • the function contains no instruction that causes an exception (since the fixup code triggered by the exception could jump back into the optimized region – Kprobes checks the exception tables to verify this);

  • there is no near jump to the optimized region (other than to the first byte).

• For each instruction in the optimized region, Kprobes verifies that the instruction can be executed out of line.

Preparing Detour Buffer

Next, Kprobes prepares a “detour” buffer, which contains the following instruction sequence:

• code to push the CPU’s registers (emulating a breakpoint trap)

• a call to the trampoline code which calls user’s probe handlers.

• code to restore registers

• the instructions from the optimized region

• a jump back to the original execution path.

Pre-optimization

After preparing the detour buffer, Kprobes verifies that none of the following situations exist:

• The probe has a post_handler.

• Other instructions in the optimized region are probed.

• The probe is disabled.

In any of the above cases, Kprobes won’t start optimizing the probe. Since these are temporary situations, Kprobes tries to start optimizing it again if the situation is changed.

If the kprobe can be optimized, Kprobes enqueues the kprobe to an optimizing list, and kicks the kprobe-optimizer workqueue to optimize it. If the to-be-optimized probepoint is hit before being optimized, Kprobes returns control to the original instruction path by setting the CPU’s instruction pointer to the copied code in the detour buffer – thus at least avoiding the single-step.

Optimization

The Kprobe-optimizer doesn’t insert the jump instruction immediately; rather, it calls synchronize_rcu() for safety first, because it’s possible for a CPU to be interrupted in the middle of executing the optimized region 3. As you know, synchronize_rcu() can ensure that all interruptions that were active when synchronize_rcu() was called are done, but only if CONFIG_PREEMPT=n. So, this version of kprobe optimization supports only kernels with CONFIG_PREEMPT=n 4.

After that, the Kprobe-optimizer calls stop_machine() to replace the optimized region with a jump instruction to the detour buffer, using text_poke_smp().

Unoptimization

When an optimized kprobe is unregistered, disabled, or blocked by another kprobe, it will be unoptimized. If this happens before the optimization is complete, the kprobe is just dequeued from the optimized list. If the optimization has been done, the jump is replaced with the original code (except for an int3 breakpoint in the first byte) by using text_poke_smp().

3

Please imagine that the 2nd instruction is interrupted and then the optimizer replaces the 2nd instruction with the jump address while the interrupt handler is running. When the interrupt returns to original address, there is no valid instruction, and it causes an unexpected result.

4

This optimization-safety checking may be replaced with the stop-machine method that ksplice uses for supporting a CONFIG_PREEMPT=y kernel.

NOTE for geeks: The jump optimization changes the kprobe’s pre_handler behavior. Without optimization, the pre_handler can change the kernel’s execution path by changing regs->ip and returning 1. However, when the probe is optimized, that modification is ignored. Thus, if you want to tweak the kernel’s execution path, you need to suppress optimization, using one of the following techniques:

• Specify an empty function for the kprobe’s post_handler.

or

• Execute ‘sysctl -w debug.kprobes_optimization=n’

Blacklist

Kprobes can probe most of the kernel except itself. This means that there are some functions where kprobes cannot probe. Probing (trapping) such functions can cause a recursive trap (e.g. double fault) or the nested probe handler may never be called. Kprobes manages such functions as a blacklist. If you want to add a function into the blacklist, you just need to (1) include linux/kprobes.h and (2) use NOKPROBE_SYMBOL() macro to specify a blacklisted function. Kprobes checks the given probe address against the blacklist and rejects registering it, if the given address is in the blacklist.

Architectures Supported

Kprobes and return probes are implemented on the following architectures:

• i386 (Supports jump optimization)

• x86_64 (AMD-64, EM64T) (Supports jump optimization)

• ppc64

• ia64 (Does not support probes on instruction slot1.)

• sparc64 (Return probes not yet implemented.)

• arm

• ppc

• mips

• s390

• parisc

Configuring Kprobes

When configuring the kernel using make menuconfig/xconfig/oldconfig, ensure that CONFIG_KPROBES is set to “y”. Under “General setup”, look for “Kprobes”.

So that you can load and unload Kprobes-based instrumentation modules, make sure “Loadable module support” (CONFIG_MODULES) and “Module unloading” (CONFIG_MODULE_UNLOAD) are set to “y”.

Also make sure that CONFIG_KALLSYMS and perhaps even CONFIG_KALLSYMS_ALL are set to “y”, since kallsyms_lookup_name() is used by the in-kernel kprobe address resolution code.

If you need to insert a probe in the middle of a function, you may find it useful to “Compile the kernel with debug info” (CONFIG_DEBUG_INFO), so you can use “objdump -d -l vmlinux” to see the source-to-object code mapping.

API Reference

The Kprobes API includes a “register” function and an “unregister” function for each type of probe. The API also includes “register_*probes” and “unregister_*probes” functions for (un)registering arrays of probes. Here are terse, mini-man-page specifications for these functions and the associated probe handlers that you’ll write. See the files in the samples/kprobes/ sub-directory for examples.

register_kprobe

#include <linux/kprobes.h>
int register_kprobe(struct kprobe *kp);

Sets a breakpoint at the address kp->addr. When the breakpoint is hit, Kprobes calls kp->pre_handler. After the probed instruction is single-stepped, Kprobe calls kp->post_handler. Any or all handlers can be NULL. If kp->flags is set KPROBE_FLAG_DISABLED, that kp will be registered but disabled, so, its handlers aren’t hit until calling enable_kprobe(kp).

Note

1. With the introduction of the “symbol_name” field to struct kprobe, the probepoint address resolution will now be taken care of by the kernel. The following will now work:

   kp.symbol_name = "symbol_name";
   

   (64-bit powerpc intricacies such as function descriptors are handled transparently)

2. Use the “offset” field of struct kprobe if the offset into the symbol to install a probepoint is known. This field is used to calculate the probepoint.

3. Specify either the kprobe “symbol_name” OR the “addr”. If both are specified, kprobe registration will fail with -EINVAL.

4. With CISC architectures (such as i386 and x86_64), the kprobes code does not validate if the kprobe.addr is at an instruction boundary. Use “offset” with caution.

register_kprobe() returns 0 on success, or a negative errno otherwise.

User’s pre-handler (kp->pre_handler):

#include <linux/kprobes.h>
#include <linux/ptrace.h>
int pre_handler(struct kprobe *p, struct pt_regs *regs);

Called with p pointing to the kprobe associated with the breakpoint, and regs pointing to the struct containing the registers saved when the breakpoint was hit. Return 0 here unless you’re a Kprobes geek.

User’s post-handler (kp->post_handler):

#include <linux/kprobes.h>
#include <linux/ptrace.h>
void post_handler(struct kprobe *p, struct pt_regs *regs,
                  unsigned long flags);

p and regs are as described for the pre_handler. flags always seems to be zero.

register_kretprobe

#include <linux/kprobes.h>
int register_kretprobe(struct kretprobe *rp);

Establishes a return probe for the function whose address is rp->kp.addr. When that function returns, Kprobes calls rp->handler. You must set rp->maxactive appropriately before you call register_kretprobe(); see “How Does a Return Probe Work?” for details.

register_kretprobe() returns 0 on success, or a negative errno otherwise.

User’s return-probe handler (rp->handler):

#include <linux/kprobes.h>
#include <linux/ptrace.h>
int kretprobe_handler(struct kretprobe_instance *ri,
                      struct pt_regs *regs);

regs is as described for kprobe.pre_handler. ri points to the kretprobe_instance object, of which the following fields may be of interest:

• ret_addr: the return address

• rp: points to the corresponding kretprobe object

• task: points to the corresponding task struct

• data: points to per return-instance private data; see “Kretprobe

  entry-handler” for details.

The regs_return_value(regs) macro provides a simple abstraction to extract the return value from the appropriate register as defined by the architecture’s ABI.

The handler’s return value is currently ignored.

unregister_*probe

#include <linux/kprobes.h>
void unregister_kprobe(struct kprobe *kp);
void unregister_kretprobe(struct kretprobe *rp);

Removes the specified probe. The unregister function can be called at any time after the probe has been registered.

Note

If the functions find an incorrect probe (ex. an unregistered probe), they clear the addr field of the probe.

register_*probes

#include <linux/kprobes.h>
int register_kprobes(struct kprobe **kps, int num);
int register_kretprobes(struct kretprobe **rps, int num);

Registers each of the num probes in the specified array. If any error occurs during registration, all probes in the array, up to the bad probe, are safely unregistered before the register_*probes function returns.

• kps/rps: an array of pointers to *probe data structures

• num: the number of the array entries.

Note

You have to allocate(or define) an array of pointers and set all of the array entries before using these functions.

unregister_*probes

#include <linux/kprobes.h>
void unregister_kprobes(struct kprobe **kps, int num);
void unregister_kretprobes(struct kretprobe **rps, int num);

Removes each of the num probes in the specified array at once.

Note

If the functions find some incorrect probes (ex. unregistered probes) in the specified array, they clear the addr field of those incorrect probes. However, other probes in the array are unregistered correctly.

disable_*probe

#include <linux/kprobes.h>
int disable_kprobe(struct kprobe *kp);
int disable_kretprobe(struct kretprobe *rp);

Temporarily disables the specified *probe. You can enable it again by using enable_*probe(). You must specify the probe which has been registered.

enable_*probe

#include <linux/kprobes.h>
int enable_kprobe(struct kprobe *kp);
int enable_kretprobe(struct kretprobe *rp);

Enables *probe which has been disabled by disable_*probe(). You must specify the probe which has been registered.

Kprobes Features and Limitations

Kprobes allows multiple probes at the same address. Also, a probepoint for which there is a post_handler cannot be optimized. So if you install a kprobe with a post_handler, at an optimized probepoint, the probepoint will be unoptimized automatically.

In general, you can install a probe anywhere in the kernel. In particular, you can probe interrupt handlers. Known exceptions are discussed in this section.

The register_*probe functions will return -EINVAL if you attempt to install a probe in the code that implements Kprobes (mostly kernel/kprobes.c and arch/*/kernel/kprobes.c, but also functions such as do_page_fault and notifier_call_chain).

If you install a probe in an inline-able function, Kprobes makes no attempt to chase down all inline instances of the function and install probes there. gcc may inline a function without being asked, so keep this in mind if you’re not seeing the probe hits you expect.

A probe handler can modify the environment of the probed function – e.g., by modifying kernel data structures, or by modifying the contents of the pt_regs struct (which are restored to the registers upon return from the breakpoint). So Kprobes can be used, for example, to install a bug fix or to inject faults for testing. Kprobes, of course, has no way to distinguish the deliberately injected faults from the accidental ones. Don’t drink and probe.

Kprobes makes no attempt to prevent probe handlers from stepping on each other – e.g., probing printk() and then calling printk() from a probe handler. If a probe handler hits a probe, that second probe’s handlers won’t be run in that instance, and the kprobe.nmissed member of the second probe will be incremented.

As of Linux v2.6.15-rc1, multiple handlers (or multiple instances of the same handler) may run concurrently on different CPUs.

Kprobes does not use mutexes or allocate memory except during registration and unregistration.

Probe handlers are run with preemption disabled or interrupt disabled, which depends on the architecture and optimization state. (e.g., kretprobe handlers and optimized kprobe handlers run without interrupt disabled on x86/x86-64). In any case, your handler should not yield the CPU (e.g., by attempting to acquire a semaphore, or waiting I/O).

Since a return probe is implemented by replacing the return address with the trampoline’s address, stack backtraces and calls to __builtin_return_address() will typically yield the trampoline’s address instead of the real return address for kretprobed functions. (As far as we can tell, __builtin_return_address() is used only for instrumentation and error reporting.)

If the number of times a function is called does not match the number of times it returns, registering a return probe on that function may produce undesirable results. In such a case, a line: kretprobe BUG!: Processing kretprobe d000000000041aa8 @ c00000000004f48c gets printed. With this information, one will be able to correlate the exact instance of the kretprobe that caused the problem. We have the do_exit() case covered. do_execve() and do_fork() are not an issue. We’re unaware of other specific cases where this could be a problem.

If, upon entry to or exit from a function, the CPU is running on a stack other than that of the current task, registering a return probe on that function may produce undesirable results. For this reason, Kprobes doesn’t support return probes (or kprobes) on the x86_64 version of __switch_to(); the registration functions return -EINVAL.

On x86/x86-64, since the Jump Optimization of Kprobes modifies instructions widely, there are some limitations to optimization. To explain it, we introduce some terminology. Imagine a 3-instruction sequence consisting of a two 2-byte instructions and one 3-byte instruction.

        IA
        |
[-2][-1][0][1][2][3][4][5][6][7]
        [ins1][ins2][  ins3 ]
        [<-     DCR       ->]
        [<- JTPR ->]

ins1: 1st Instruction
ins2: 2nd Instruction
ins3: 3rd Instruction
IA:  Insertion Address
JTPR: Jump Target Prohibition Region
DCR: Detoured Code Region

The instructions in DCR are copied to the out-of-line buffer of the kprobe, because the bytes in DCR are replaced by a 5-byte jump instruction. So there are several limitations.

1. The instructions in DCR must be relocatable.

2. The instructions in DCR must not include a call instruction.

3. JTPR must not be targeted by any jump or call instruction.

4. DCR must not straddle the border between functions.

Anyway, these limitations are checked by the in-kernel instruction decoder, so you don’t need to worry about that.

Probe Overhead

On a typical CPU in use in 2005, a kprobe hit takes 0.5 to 1.0 microseconds to process. Specifically, a benchmark that hits the same probepoint repeatedly, firing a simple handler each time, reports 1-2 million hits per second, depending on the architecture. A return-probe hit typically takes 50-75% longer than a kprobe hit. When you have a return probe set on a function, adding a kprobe at the entry to that function adds essentially no overhead.

Here are sample overhead figures (in usec) for different architectures:

k = kprobe; r = return probe; kr = kprobe + return probe
on same function

i386: Intel Pentium M, 1495 MHz, 2957.31 bogomips
k = 0.57 usec; r = 0.92; kr = 0.99

x86_64: AMD Opteron 246, 1994 MHz, 3971.48 bogomips
k = 0.49 usec; r = 0.80; kr = 0.82

ppc64: POWER5 (gr), 1656 MHz (SMT disabled, 1 virtual CPU per physical CPU)
k = 0.77 usec; r = 1.26; kr = 1.45

Optimized Probe Overhead

Typically, an optimized kprobe hit takes 0.07 to 0.1 microseconds to process. Here are sample overhead figures (in usec) for x86 architectures:

k = unoptimized kprobe, b = boosted (single-step skipped), o = optimized kprobe,
r = unoptimized kretprobe, rb = boosted kretprobe, ro = optimized kretprobe.

i386: Intel(R) Xeon(R) E5410, 2.33GHz, 4656.90 bogomips
k = 0.80 usec; b = 0.33; o = 0.05; r = 1.10; rb = 0.61; ro = 0.33

x86-64: Intel(R) Xeon(R) E5410, 2.33GHz, 4656.90 bogomips
k = 0.99 usec; b = 0.43; o = 0.06; r = 1.24; rb = 0.68; ro = 0.30

TODO

1. SystemTap (http://sourceware.org/systemtap): Provides a simplified programming interface for probe-based instrumentation. Try it out.

2. Kernel return probes for sparc64.

3. Support for other architectures.

4. User-space probes.

5. Watchpoint probes (which fire on data references).

Kprobes Example

See samples/kprobes/kprobe_example.c

Kretprobes Example

See samples/kprobes/kretprobe_example.c

Deprecated Features

Jprobes is now a deprecated feature. People who are depending on it should migrate to other tracing features or use older kernels. Please consider to migrate your tool to one of the following options:

• Use trace-event to trace target function with arguments.

  trace-event is a low-overhead (and almost no visible overhead if it is off) statically defined event interface. You can define new events and trace it via ftrace or any other tracing tools.

  See the following urls:

  • https://lwn.net/Articles/379903/

  • https://lwn.net/Articles/381064/

  • https://lwn.net/Articles/383362/

• Use ftrace dynamic events (kprobe event) with perf-probe.

  If you build your kernel with debug info (CONFIG_DEBUG_INFO=y), you can find which register/stack is assigned to which local variable or arguments by using perf-probe and set up new event to trace it.

  See following documents:

  • Kprobe-based Event Tracing

  • Event Tracing

  • tools/perf/Documentation/perf-probe.txt

The kprobes debugfs interface

With recent kernels (> 2.6.20) the list of registered kprobes is visible under the /sys/kernel/debug/kprobes/ directory (assuming debugfs is mounted at //sys/kernel/debug).

/sys/kernel/debug/kprobes/list: Lists all registered probes on the system:

c015d71a  k  vfs_read+0x0
c03dedc5  r  tcp_v4_rcv+0x0

The first column provides the kernel address where the probe is inserted. The second column identifies the type of probe (k - kprobe and r - kretprobe) while the third column specifies the symbol+offset of the probe. If the probed function belongs to a module, the module name is also specified. Following columns show probe status. If the probe is on a virtual address that is no longer valid (module init sections, module virtual addresses that correspond to modules that’ve been unloaded), such probes are marked with [GONE]. If the probe is temporarily disabled, such probes are marked with [DISABLED]. If the probe is optimized, it is marked with [OPTIMIZED]. If the probe is ftrace-based, it is marked with [FTRACE].

/sys/kernel/debug/kprobes/enabled: Turn kprobes ON/OFF forcibly.

Provides a knob to globally and forcibly turn registered kprobes ON or OFF. By default, all kprobes are enabled. By echoing “0” to this file, all registered probes will be disarmed, till such time a “1” is echoed to this file. Note that this knob just disarms and arms all kprobes and doesn’t change each probe’s disabling state. This means that disabled kprobes (marked [DISABLED]) will be not enabled if you turn ON all kprobes by this knob.

The kprobes sysctl interface

/proc/sys/debug/kprobes-optimization: Turn kprobes optimization ON/OFF.

When CONFIG_OPTPROBES=y, this sysctl interface appears and it provides a knob to globally and forcibly turn jump optimization (see section How Does Jump Optimization Work?) ON or OFF. By default, jump optimization is allowed (ON). If you echo “0” to this file or set “debug.kprobes_optimization” to 0 via sysctl, all optimized probes will be unoptimized, and any new probes registered after that will not be optimized.

Note that this knob changes the optimized state. This means that optimized probes (marked [OPTIMIZED]) will be unoptimized ([OPTIMIZED] tag will be removed). If the knob is turned on, they will be optimized again.

References

For additional information on Kprobes, refer to the following URLs:

• https://lwn.net/Articles/132196/

• https://www.kernel.org/doc/ols/2006/ols2006v2-pages-109-124.pdf