5 This document outlines basic information about kernel livepatching.
10 2. Kprobes, Ftrace, Livepatching
15 4.3. Livepatch module handling
16 5. Livepatch life-cycle
28 There are many situations where users are reluctant to reboot a system. It may
29 be because their system is performing complex scientific computations or under
30 heavy load during peak usage. In addition to keeping systems up and running,
31 users want to also have a stable and secure system. Livepatching gives users
32 both by allowing for function calls to be redirected; thus, fixing critical
33 functions without a system reboot.
36 2. Kprobes, Ftrace, Livepatching
37 ================================
39 There are multiple mechanisms in the Linux kernel that are directly related
40 to redirection of code execution; namely: kernel probes, function tracing,
43 + The kernel probes are the most generic. The code can be redirected by
44 putting a breakpoint instruction instead of any instruction.
46 + The function tracer calls the code from a predefined location that is
47 close to the function entry point. This location is generated by the
48 compiler using the '-pg' gcc option.
50 + Livepatching typically needs to redirect the code at the very beginning
51 of the function entry before the function parameters or the stack
52 are in any way modified.
54 All three approaches need to modify the existing code at runtime. Therefore
55 they need to be aware of each other and not step over each other's toes.
56 Most of these problems are solved by using the dynamic ftrace framework as
57 a base. A Kprobe is registered as a ftrace handler when the function entry
58 is probed, see CONFIG_KPROBES_ON_FTRACE. Also an alternative function from
59 a live patch is called with the help of a custom ftrace handler. But there are
60 some limitations, see below.
66 Functions are there for a reason. They take some input parameters, get or
67 release locks, read, process, and even write some data in a defined way,
68 have return values. In other words, each function has a defined semantic.
70 Many fixes do not change the semantic of the modified functions. For
71 example, they add a NULL pointer or a boundary check, fix a race by adding
72 a missing memory barrier, or add some locking around a critical section.
73 Most of these changes are self contained and the function presents itself
74 the same way to the rest of the system. In this case, the functions might
75 be updated independently one by one.
77 But there are more complex fixes. For example, a patch might change
78 ordering of locking in multiple functions at the same time. Or a patch
79 might exchange meaning of some temporary structures and update
80 all the relevant functions. In this case, the affected unit
81 (thread, whole kernel) need to start using all new versions of
82 the functions at the same time. Also the switch must happen only
83 when it is safe to do so, e.g. when the affected locks are released
84 or no data are stored in the modified structures at the moment.
86 The theory about how to apply functions a safe way is rather complex.
87 The aim is to define a so-called consistency model. It attempts to define
88 conditions when the new implementation could be used so that the system
91 Livepatch has a consistency model which is a hybrid of kGraft and
92 kpatch: it uses kGraft's per-task consistency and syscall barrier
93 switching combined with kpatch's stack trace switching. There are also
94 a number of fallback options which make it quite flexible.
96 Patches are applied on a per-task basis, when the task is deemed safe to
97 switch over. When a patch is enabled, livepatch enters into a
98 transition state where tasks are converging to the patched state.
99 Usually this transition state can complete in a few seconds. The same
100 sequence occurs when a patch is disabled, except the tasks converge from
101 the patched state to the unpatched state.
103 An interrupt handler inherits the patched state of the task it
104 interrupts. The same is true for forked tasks: the child inherits the
105 patched state of the parent.
107 Livepatch uses several complementary approaches to determine when it's
110 1. The first and most effective approach is stack checking of sleeping
111 tasks. If no affected functions are on the stack of a given task,
112 the task is patched. In most cases this will patch most or all of
113 the tasks on the first try. Otherwise it'll keep trying
114 periodically. This option is only available if the architecture has
115 reliable stacks (HAVE_RELIABLE_STACKTRACE).
117 2. The second approach, if needed, is kernel exit switching. A
118 task is switched when it returns to user space from a system call, a
119 user space IRQ, or a signal. It's useful in the following cases:
121 a) Patching I/O-bound user tasks which are sleeping on an affected
122 function. In this case you have to send SIGSTOP and SIGCONT to
123 force it to exit the kernel and be patched.
124 b) Patching CPU-bound user tasks. If the task is highly CPU-bound
125 then it will get patched the next time it gets interrupted by an
128 3. For idle "swapper" tasks, since they don't ever exit the kernel, they
129 instead have a klp_update_patch_state() call in the idle loop which
130 allows them to be patched before the CPU enters the idle state.
132 (Note there's not yet such an approach for kthreads.)
134 Architectures which don't have HAVE_RELIABLE_STACKTRACE solely rely on
135 the second approach. It's highly likely that some tasks may still be
136 running with an old version of the function, until that function
137 returns. In this case you would have to signal the tasks. This
138 especially applies to kthreads. They may not be woken up and would need
139 to be forced. See below for more information.
141 Unless we can come up with another way to patch kthreads, architectures
142 without HAVE_RELIABLE_STACKTRACE are not considered fully supported by
143 the kernel livepatching.
145 The /sys/kernel/livepatch/<patch>/transition file shows whether a patch
146 is in transition. Only a single patch (the topmost patch on the stack)
147 can be in transition at a given time. A patch can remain in transition
148 indefinitely, if any of the tasks are stuck in the initial patch state.
150 A transition can be reversed and effectively canceled by writing the
151 opposite value to the /sys/kernel/livepatch/<patch>/enabled file while
152 the transition is in progress. Then all the tasks will attempt to
153 converge back to the original patch state.
155 There's also a /proc/<pid>/patch_state file which can be used to
156 determine which tasks are blocking completion of a patching operation.
157 If a patch is in transition, this file shows 0 to indicate the task is
158 unpatched and 1 to indicate it's patched. Otherwise, if no patch is in
159 transition, it shows -1. Any tasks which are blocking the transition
160 can be signaled with SIGSTOP and SIGCONT to force them to change their
161 patched state. This may be harmful to the system though.
162 /sys/kernel/livepatch/<patch>/signal attribute provides a better alternative.
163 Writing 1 to the attribute sends a fake signal to all remaining blocking
164 tasks. No proper signal is actually delivered (there is no data in signal
165 pending structures). Tasks are interrupted or woken up, and forced to change
168 Administrator can also affect a transition through
169 /sys/kernel/livepatch/<patch>/force attribute. Writing 1 there clears
170 TIF_PATCH_PENDING flag of all tasks and thus forces the tasks to the patched
171 state. Important note! The force attribute is intended for cases when the
172 transition gets stuck for a long time because of a blocking task. Administrator
173 is expected to collect all necessary data (namely stack traces of such blocking
174 tasks) and request a clearance from a patch distributor to force the transition.
175 Unauthorized usage may cause harm to the system. It depends on the nature of the
176 patch, which functions are (un)patched, and which functions the blocking tasks
177 are sleeping in (/proc/<pid>/stack may help here). Removal (rmmod) of patch
178 modules is permanently disabled when the force feature is used. It cannot be
179 guaranteed there is no task sleeping in such module. It implies unbounded
180 reference count if a patch module is disabled and enabled in a loop.
182 Moreover, the usage of force may also affect future applications of live
183 patches and cause even more harm to the system. Administrator should first
184 consider to simply cancel a transition (see above). If force is used, reboot
185 should be planned and no more live patches applied.
187 3.1 Adding consistency model support to new architectures
188 ---------------------------------------------------------
190 For adding consistency model support to new architectures, there are a
193 1) Add CONFIG_HAVE_RELIABLE_STACKTRACE. This means porting objtool, and
194 for non-DWARF unwinders, also making sure there's a way for the stack
195 tracing code to detect interrupts on the stack.
197 2) Alternatively, ensure that every kthread has a call to
198 klp_update_patch_state() in a safe location. Kthreads are typically
199 in an infinite loop which does some action repeatedly. The safe
200 location to switch the kthread's patch state would be at a designated
201 point in the loop where there are no locks taken and all data
202 structures are in a well-defined state.
204 The location is clear when using workqueues or the kthread worker
205 API. These kthreads process independent actions in a generic loop.
207 It's much more complicated with kthreads which have a custom loop.
208 There the safe location must be carefully selected on a case-by-case
211 In that case, arches without HAVE_RELIABLE_STACKTRACE would still be
212 able to use the non-stack-checking parts of the consistency model:
214 a) patching user tasks when they cross the kernel/user space
217 b) patching kthreads and idle tasks at their designated patch points.
219 This option isn't as good as option 1 because it requires signaling
220 user tasks and waking kthreads to patch them. But it could still be
221 a good backup option for those architectures which don't have
222 reliable stack traces yet.
228 Livepatches are distributed using kernel modules, see
229 samples/livepatch/livepatch-sample.c.
231 The module includes a new implementation of functions that we want
232 to replace. In addition, it defines some structures describing the
233 relation between the original and the new implementation. Then there
234 is code that makes the kernel start using the new code when the livepatch
235 module is loaded. Also there is code that cleans up before the
236 livepatch module is removed. All this is explained in more details in
243 New versions of functions are typically just copied from the original
244 sources. A good practice is to add a prefix to the names so that they
245 can be distinguished from the original ones, e.g. in a backtrace. Also
246 they can be declared as static because they are not called directly
247 and do not need the global visibility.
249 The patch contains only functions that are really modified. But they
250 might want to access functions or data from the original source file
251 that may only be locally accessible. This can be solved by a special
252 relocation section in the generated livepatch module, see
253 Documentation/livepatch/module-elf-format.txt for more details.
259 The patch is described by several structures that split the information
262 + struct klp_func is defined for each patched function. It describes
263 the relation between the original and the new implementation of a
266 The structure includes the name, as a string, of the original function.
267 The function address is found via kallsyms at runtime.
269 Then it includes the address of the new function. It is defined
270 directly by assigning the function pointer. Note that the new
271 function is typically defined in the same source file.
273 As an optional parameter, the symbol position in the kallsyms database can
274 be used to disambiguate functions of the same name. This is not the
275 absolute position in the database, but rather the order it has been found
276 only for a particular object ( vmlinux or a kernel module ). Note that
277 kallsyms allows for searching symbols according to the object name.
279 + struct klp_object defines an array of patched functions (struct
280 klp_func) in the same object. Where the object is either vmlinux
281 (NULL) or a module name.
283 The structure helps to group and handle functions for each object
284 together. Note that patched modules might be loaded later than
285 the patch itself and the relevant functions might be patched
286 only when they are available.
289 + struct klp_patch defines an array of patched objects (struct
292 This structure handles all patched functions consistently and eventually,
293 synchronously. The whole patch is applied only when all patched
294 symbols are found. The only exception are symbols from objects
295 (kernel modules) that have not been loaded yet.
297 For more details on how the patch is applied on a per-task basis,
298 see the "Consistency model" section.
301 4.3. Livepatch module handling
302 ------------------------------
304 The usual behavior is that the new functions will get used when
305 the livepatch module is loaded. For this, the module init() function
306 has to register the patch (struct klp_patch) and enable it. See the
307 section "Livepatch life-cycle" below for more details about these
310 Module removal is only safe when there are no users of the underlying
311 functions. This is the reason why the force feature permanently disables
312 the removal. The forced tasks entered the functions but we cannot say
313 that they returned back. Therefore it cannot be decided when the
314 livepatch module can be safely removed. When the system is successfully
315 transitioned to a new patch state (patched/unpatched) without being
316 forced it is guaranteed that no task sleeps or runs in the old code.
319 5. Livepatch life-cycle
320 =======================
322 Livepatching defines four basic operations that define the life cycle of each
323 live patch: registration, enabling, disabling and unregistration. There are
324 several reasons why it is done this way.
326 First, the patch is applied only when all patched symbols for already
327 loaded objects are found. The error handling is much easier if this
328 check is done before particular functions get redirected.
330 Second, it might take some time until the entire system is migrated with
331 the hybrid consistency model being used. The patch revert might block
332 the livepatch module removal for too long. Therefore it is useful to
333 revert the patch using a separate operation that might be called
334 explicitly. But it does not make sense to remove all information until
335 the livepatch module is really removed.
341 Each patch first has to be registered using klp_register_patch(). This makes
342 the patch known to the livepatch framework. Also it does some preliminary
343 computing and checks.
345 In particular, the patch is added into the list of known patches. The
346 addresses of the patched functions are found according to their names.
347 The special relocations, mentioned in the section "New functions", are
348 applied. The relevant entries are created under
349 /sys/kernel/livepatch/<name>. The patch is rejected when any operation
356 Registered patches might be enabled either by calling klp_enable_patch() or
357 by writing '1' to /sys/kernel/livepatch/<name>/enabled. The system will
358 start using the new implementation of the patched functions at this stage.
360 When a patch is enabled, livepatch enters into a transition state where
361 tasks are converging to the patched state. This is indicated by a value
362 of '1' in /sys/kernel/livepatch/<name>/transition. Once all tasks have
363 been patched, the 'transition' value changes to '0'. For more
364 information about this process, see the "Consistency model" section.
366 If an original function is patched for the first time, a function
367 specific struct klp_ops is created and an universal ftrace handler is
370 Functions might be patched multiple times. The ftrace handler is registered
371 only once for the given function. Further patches just add an entry to the
372 list (see field `func_stack`) of the struct klp_ops. The last added
373 entry is chosen by the ftrace handler and becomes the active function
376 Note that the patches might be enabled in a different order than they were
383 Enabled patches might get disabled either by calling klp_disable_patch() or
384 by writing '0' to /sys/kernel/livepatch/<name>/enabled. At this stage
385 either the code from the previously enabled patch or even the original
388 When a patch is disabled, livepatch enters into a transition state where
389 tasks are converging to the unpatched state. This is indicated by a
390 value of '1' in /sys/kernel/livepatch/<name>/transition. Once all tasks
391 have been unpatched, the 'transition' value changes to '0'. For more
392 information about this process, see the "Consistency model" section.
394 Here all the functions (struct klp_func) associated with the to-be-disabled
395 patch are removed from the corresponding struct klp_ops. The ftrace handler
396 is unregistered and the struct klp_ops is freed when the func_stack list
399 Patches must be disabled in exactly the reverse order in which they were
400 enabled. It makes the problem and the implementation much easier.
406 Disabled patches might be unregistered by calling klp_unregister_patch().
407 This can be done only when the patch is disabled and the code is no longer
408 used. It must be called before the livepatch module gets unloaded.
410 At this stage, all the relevant sys-fs entries are removed and the patch
411 is removed from the list of known patches.
417 Information about the registered patches can be found under
418 /sys/kernel/livepatch. The patches could be enabled and disabled
421 /sys/kernel/livepatch/<patch>/signal and /sys/kernel/livepatch/<patch>/force
422 attributes allow administrator to affect a patching operation.
424 See Documentation/ABI/testing/sysfs-kernel-livepatch for more details.
430 The current Livepatch implementation has several limitations:
433 + The patch must not change the semantic of the patched functions.
435 The current implementation guarantees only that either the old
436 or the new function is called. The functions are patched one
437 by one. It means that the patch must _not_ change the semantic
441 + Data structures can not be patched.
443 There is no support to version data structures or anyhow migrate
444 one structure into another. Also the simple consistency model does
445 not allow to switch more functions atomically.
447 Once there is more complex consistency mode, it will be possible to
448 use some workarounds. For example, it will be possible to use a hole
449 for a new member because the data structure is aligned. Or it will
450 be possible to use an existing member for something else.
452 There are no plans to add more generic support for modified structures
456 + Only functions that can be traced could be patched.
458 Livepatch is based on the dynamic ftrace. In particular, functions
459 implementing ftrace or the livepatch ftrace handler could not be
460 patched. Otherwise, the code would end up in an infinite loop. A
461 potential mistake is prevented by marking the problematic functions
466 + Livepatch works reliably only when the dynamic ftrace is located at
467 the very beginning of the function.
469 The function need to be redirected before the stack or the function
470 parameters are modified in any way. For example, livepatch requires
471 using -fentry gcc compiler option on x86_64.
473 One exception is the PPC port. It uses relative addressing and TOC.
474 Each function has to handle TOC and save LR before it could call
475 the ftrace handler. This operation has to be reverted on return.
476 Fortunately, the generic ftrace code has the same problem and all
477 this is handled on the ftrace level.
480 + Kretprobes using the ftrace framework conflict with the patched
483 Both kretprobes and livepatches use a ftrace handler that modifies
484 the return address. The first user wins. Either the probe or the patch
485 is rejected when the handler is already in use by the other.
488 + Kprobes in the original function are ignored when the code is
489 redirected to the new implementation.
491 There is a work in progress to add warnings about this situation.