livepatch.rst 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461
  1. =========
  2. Livepatch
  3. =========
  4. This document outlines basic information about kernel livepatching.
  5. .. Table of Contents:
  6. 1. Motivation
  7. 2. Kprobes, Ftrace, Livepatching
  8. 3. Consistency model
  9. 4. Livepatch module
  10. 4.1. New functions
  11. 4.2. Metadata
  12. 5. Livepatch life-cycle
  13. 5.1. Loading
  14. 5.2. Enabling
  15. 5.3. Replacing
  16. 5.4. Disabling
  17. 5.5. Removing
  18. 6. Sysfs
  19. 7. Limitations
  20. 1. Motivation
  21. =============
  22. There are many situations where users are reluctant to reboot a system. It may
  23. be because their system is performing complex scientific computations or under
  24. heavy load during peak usage. In addition to keeping systems up and running,
  25. users want to also have a stable and secure system. Livepatching gives users
  26. both by allowing for function calls to be redirected; thus, fixing critical
  27. functions without a system reboot.
  28. 2. Kprobes, Ftrace, Livepatching
  29. ================================
  30. There are multiple mechanisms in the Linux kernel that are directly related
  31. to redirection of code execution; namely: kernel probes, function tracing,
  32. and livepatching:
  33. - The kernel probes are the most generic. The code can be redirected by
  34. putting a breakpoint instruction instead of any instruction.
  35. - The function tracer calls the code from a predefined location that is
  36. close to the function entry point. This location is generated by the
  37. compiler using the '-pg' gcc option.
  38. - Livepatching typically needs to redirect the code at the very beginning
  39. of the function entry before the function parameters or the stack
  40. are in any way modified.
  41. All three approaches need to modify the existing code at runtime. Therefore
  42. they need to be aware of each other and not step over each other's toes.
  43. Most of these problems are solved by using the dynamic ftrace framework as
  44. a base. A Kprobe is registered as a ftrace handler when the function entry
  45. is probed, see CONFIG_KPROBES_ON_FTRACE. Also an alternative function from
  46. a live patch is called with the help of a custom ftrace handler. But there are
  47. some limitations, see below.
  48. 3. Consistency model
  49. ====================
  50. Functions are there for a reason. They take some input parameters, get or
  51. release locks, read, process, and even write some data in a defined way,
  52. have return values. In other words, each function has a defined semantic.
  53. Many fixes do not change the semantic of the modified functions. For
  54. example, they add a NULL pointer or a boundary check, fix a race by adding
  55. a missing memory barrier, or add some locking around a critical section.
  56. Most of these changes are self contained and the function presents itself
  57. the same way to the rest of the system. In this case, the functions might
  58. be updated independently one by one.
  59. But there are more complex fixes. For example, a patch might change
  60. ordering of locking in multiple functions at the same time. Or a patch
  61. might exchange meaning of some temporary structures and update
  62. all the relevant functions. In this case, the affected unit
  63. (thread, whole kernel) need to start using all new versions of
  64. the functions at the same time. Also the switch must happen only
  65. when it is safe to do so, e.g. when the affected locks are released
  66. or no data are stored in the modified structures at the moment.
  67. The theory about how to apply functions a safe way is rather complex.
  68. The aim is to define a so-called consistency model. It attempts to define
  69. conditions when the new implementation could be used so that the system
  70. stays consistent.
  71. Livepatch has a consistency model which is a hybrid of kGraft and
  72. kpatch: it uses kGraft's per-task consistency and syscall barrier
  73. switching combined with kpatch's stack trace switching. There are also
  74. a number of fallback options which make it quite flexible.
  75. Patches are applied on a per-task basis, when the task is deemed safe to
  76. switch over. When a patch is enabled, livepatch enters into a
  77. transition state where tasks are converging to the patched state.
  78. Usually this transition state can complete in a few seconds. The same
  79. sequence occurs when a patch is disabled, except the tasks converge from
  80. the patched state to the unpatched state.
  81. An interrupt handler inherits the patched state of the task it
  82. interrupts. The same is true for forked tasks: the child inherits the
  83. patched state of the parent.
  84. Livepatch uses several complementary approaches to determine when it's
  85. safe to patch tasks:
  86. 1. The first and most effective approach is stack checking of sleeping
  87. tasks. If no affected functions are on the stack of a given task,
  88. the task is patched. In most cases this will patch most or all of
  89. the tasks on the first try. Otherwise it'll keep trying
  90. periodically. This option is only available if the architecture has
  91. reliable stacks (HAVE_RELIABLE_STACKTRACE).
  92. 2. The second approach, if needed, is kernel exit switching. A
  93. task is switched when it returns to user space from a system call, a
  94. user space IRQ, or a signal. It's useful in the following cases:
  95. a) Patching I/O-bound user tasks which are sleeping on an affected
  96. function. In this case you have to send SIGSTOP and SIGCONT to
  97. force it to exit the kernel and be patched.
  98. b) Patching CPU-bound user tasks. If the task is highly CPU-bound
  99. then it will get patched the next time it gets interrupted by an
  100. IRQ.
  101. 3. For idle "swapper" tasks, since they don't ever exit the kernel, they
  102. instead have a klp_update_patch_state() call in the idle loop which
  103. allows them to be patched before the CPU enters the idle state.
  104. (Note there's not yet such an approach for kthreads.)
  105. Architectures which don't have HAVE_RELIABLE_STACKTRACE solely rely on
  106. the second approach. It's highly likely that some tasks may still be
  107. running with an old version of the function, until that function
  108. returns. In this case you would have to signal the tasks. This
  109. especially applies to kthreads. They may not be woken up and would need
  110. to be forced. See below for more information.
  111. Unless we can come up with another way to patch kthreads, architectures
  112. without HAVE_RELIABLE_STACKTRACE are not considered fully supported by
  113. the kernel livepatching.
  114. The /sys/kernel/livepatch/<patch>/transition file shows whether a patch
  115. is in transition. Only a single patch can be in transition at a given
  116. time. A patch can remain in transition indefinitely, if any of the tasks
  117. are stuck in the initial patch state.
  118. A transition can be reversed and effectively canceled by writing the
  119. opposite value to the /sys/kernel/livepatch/<patch>/enabled file while
  120. the transition is in progress. Then all the tasks will attempt to
  121. converge back to the original patch state.
  122. There's also a /proc/<pid>/patch_state file which can be used to
  123. determine which tasks are blocking completion of a patching operation.
  124. If a patch is in transition, this file shows 0 to indicate the task is
  125. unpatched and 1 to indicate it's patched. Otherwise, if no patch is in
  126. transition, it shows -1. Any tasks which are blocking the transition
  127. can be signaled with SIGSTOP and SIGCONT to force them to change their
  128. patched state. This may be harmful to the system though. Sending a fake signal
  129. to all remaining blocking tasks is a better alternative. No proper signal is
  130. actually delivered (there is no data in signal pending structures). Tasks are
  131. interrupted or woken up, and forced to change their patched state. The fake
  132. signal is automatically sent every 15 seconds.
  133. Administrator can also affect a transition through
  134. /sys/kernel/livepatch/<patch>/force attribute. Writing 1 there clears
  135. TIF_PATCH_PENDING flag of all tasks and thus forces the tasks to the patched
  136. state. Important note! The force attribute is intended for cases when the
  137. transition gets stuck for a long time because of a blocking task. Administrator
  138. is expected to collect all necessary data (namely stack traces of such blocking
  139. tasks) and request a clearance from a patch distributor to force the transition.
  140. Unauthorized usage may cause harm to the system. It depends on the nature of the
  141. patch, which functions are (un)patched, and which functions the blocking tasks
  142. are sleeping in (/proc/<pid>/stack may help here). Removal (rmmod) of patch
  143. modules is permanently disabled when the force feature is used. It cannot be
  144. guaranteed there is no task sleeping in such module. It implies unbounded
  145. reference count if a patch module is disabled and enabled in a loop.
  146. Moreover, the usage of force may also affect future applications of live
  147. patches and cause even more harm to the system. Administrator should first
  148. consider to simply cancel a transition (see above). If force is used, reboot
  149. should be planned and no more live patches applied.
  150. 3.1 Adding consistency model support to new architectures
  151. ---------------------------------------------------------
  152. For adding consistency model support to new architectures, there are a
  153. few options:
  154. 1) Add CONFIG_HAVE_RELIABLE_STACKTRACE. This means porting objtool, and
  155. for non-DWARF unwinders, also making sure there's a way for the stack
  156. tracing code to detect interrupts on the stack.
  157. 2) Alternatively, ensure that every kthread has a call to
  158. klp_update_patch_state() in a safe location. Kthreads are typically
  159. in an infinite loop which does some action repeatedly. The safe
  160. location to switch the kthread's patch state would be at a designated
  161. point in the loop where there are no locks taken and all data
  162. structures are in a well-defined state.
  163. The location is clear when using workqueues or the kthread worker
  164. API. These kthreads process independent actions in a generic loop.
  165. It's much more complicated with kthreads which have a custom loop.
  166. There the safe location must be carefully selected on a case-by-case
  167. basis.
  168. In that case, arches without HAVE_RELIABLE_STACKTRACE would still be
  169. able to use the non-stack-checking parts of the consistency model:
  170. a) patching user tasks when they cross the kernel/user space
  171. boundary; and
  172. b) patching kthreads and idle tasks at their designated patch points.
  173. This option isn't as good as option 1 because it requires signaling
  174. user tasks and waking kthreads to patch them. But it could still be
  175. a good backup option for those architectures which don't have
  176. reliable stack traces yet.
  177. 4. Livepatch module
  178. ===================
  179. Livepatches are distributed using kernel modules, see
  180. samples/livepatch/livepatch-sample.c.
  181. The module includes a new implementation of functions that we want
  182. to replace. In addition, it defines some structures describing the
  183. relation between the original and the new implementation. Then there
  184. is code that makes the kernel start using the new code when the livepatch
  185. module is loaded. Also there is code that cleans up before the
  186. livepatch module is removed. All this is explained in more details in
  187. the next sections.
  188. 4.1. New functions
  189. ------------------
  190. New versions of functions are typically just copied from the original
  191. sources. A good practice is to add a prefix to the names so that they
  192. can be distinguished from the original ones, e.g. in a backtrace. Also
  193. they can be declared as static because they are not called directly
  194. and do not need the global visibility.
  195. The patch contains only functions that are really modified. But they
  196. might want to access functions or data from the original source file
  197. that may only be locally accessible. This can be solved by a special
  198. relocation section in the generated livepatch module, see
  199. Documentation/livepatch/module-elf-format.rst for more details.
  200. 4.2. Metadata
  201. -------------
  202. The patch is described by several structures that split the information
  203. into three levels:
  204. - struct klp_func is defined for each patched function. It describes
  205. the relation between the original and the new implementation of a
  206. particular function.
  207. The structure includes the name, as a string, of the original function.
  208. The function address is found via kallsyms at runtime.
  209. Then it includes the address of the new function. It is defined
  210. directly by assigning the function pointer. Note that the new
  211. function is typically defined in the same source file.
  212. As an optional parameter, the symbol position in the kallsyms database can
  213. be used to disambiguate functions of the same name. This is not the
  214. absolute position in the database, but rather the order it has been found
  215. only for a particular object ( vmlinux or a kernel module ). Note that
  216. kallsyms allows for searching symbols according to the object name.
  217. - struct klp_object defines an array of patched functions (struct
  218. klp_func) in the same object. Where the object is either vmlinux
  219. (NULL) or a module name.
  220. The structure helps to group and handle functions for each object
  221. together. Note that patched modules might be loaded later than
  222. the patch itself and the relevant functions might be patched
  223. only when they are available.
  224. - struct klp_patch defines an array of patched objects (struct
  225. klp_object).
  226. This structure handles all patched functions consistently and eventually,
  227. synchronously. The whole patch is applied only when all patched
  228. symbols are found. The only exception are symbols from objects
  229. (kernel modules) that have not been loaded yet.
  230. For more details on how the patch is applied on a per-task basis,
  231. see the "Consistency model" section.
  232. 5. Livepatch life-cycle
  233. =======================
  234. Livepatching can be described by five basic operations:
  235. loading, enabling, replacing, disabling, removing.
  236. Where the replacing and the disabling operations are mutually
  237. exclusive. They have the same result for the given patch but
  238. not for the system.
  239. 5.1. Loading
  240. ------------
  241. The only reasonable way is to enable the patch when the livepatch kernel
  242. module is being loaded. For this, klp_enable_patch() has to be called
  243. in the module_init() callback. There are two main reasons:
  244. First, only the module has an easy access to the related struct klp_patch.
  245. Second, the error code might be used to refuse loading the module when
  246. the patch cannot get enabled.
  247. 5.2. Enabling
  248. -------------
  249. The livepatch gets enabled by calling klp_enable_patch() from
  250. the module_init() callback. The system will start using the new
  251. implementation of the patched functions at this stage.
  252. First, the addresses of the patched functions are found according to their
  253. names. The special relocations, mentioned in the section "New functions",
  254. are applied. The relevant entries are created under
  255. /sys/kernel/livepatch/<name>. The patch is rejected when any above
  256. operation fails.
  257. Second, livepatch enters into a transition state where tasks are converging
  258. to the patched state. If an original function is patched for the first
  259. time, a function specific struct klp_ops is created and an universal
  260. ftrace handler is registered\ [#]_. This stage is indicated by a value of '1'
  261. in /sys/kernel/livepatch/<name>/transition. For more information about
  262. this process, see the "Consistency model" section.
  263. Finally, once all tasks have been patched, the 'transition' value changes
  264. to '0'.
  265. .. [#]
  266. Note that functions might be patched multiple times. The ftrace handler
  267. is registered only once for a given function. Further patches just add
  268. an entry to the list (see field `func_stack`) of the struct klp_ops.
  269. The right implementation is selected by the ftrace handler, see
  270. the "Consistency model" section.
  271. That said, it is highly recommended to use cumulative livepatches
  272. because they help keeping the consistency of all changes. In this case,
  273. functions might be patched two times only during the transition period.
  274. 5.3. Replacing
  275. --------------
  276. All enabled patches might get replaced by a cumulative patch that
  277. has the .replace flag set.
  278. Once the new patch is enabled and the 'transition' finishes then
  279. all the functions (struct klp_func) associated with the replaced
  280. patches are removed from the corresponding struct klp_ops. Also
  281. the ftrace handler is unregistered and the struct klp_ops is
  282. freed when the related function is not modified by the new patch
  283. and func_stack list becomes empty.
  284. See Documentation/livepatch/cumulative-patches.rst for more details.
  285. 5.4. Disabling
  286. --------------
  287. Enabled patches might get disabled by writing '0' to
  288. /sys/kernel/livepatch/<name>/enabled.
  289. First, livepatch enters into a transition state where tasks are converging
  290. to the unpatched state. The system starts using either the code from
  291. the previously enabled patch or even the original one. This stage is
  292. indicated by a value of '1' in /sys/kernel/livepatch/<name>/transition.
  293. For more information about this process, see the "Consistency model"
  294. section.
  295. Second, once all tasks have been unpatched, the 'transition' value changes
  296. to '0'. All the functions (struct klp_func) associated with the to-be-disabled
  297. patch are removed from the corresponding struct klp_ops. The ftrace handler
  298. is unregistered and the struct klp_ops is freed when the func_stack list
  299. becomes empty.
  300. Third, the sysfs interface is destroyed.
  301. 5.5. Removing
  302. -------------
  303. Module removal is only safe when there are no users of functions provided
  304. by the module. This is the reason why the force feature permanently
  305. disables the removal. Only when the system is successfully transitioned
  306. to a new patch state (patched/unpatched) without being forced it is
  307. guaranteed that no task sleeps or runs in the old code.
  308. 6. Sysfs
  309. ========
  310. Information about the registered patches can be found under
  311. /sys/kernel/livepatch. The patches could be enabled and disabled
  312. by writing there.
  313. /sys/kernel/livepatch/<patch>/force attributes allow administrator to affect a
  314. patching operation.
  315. See Documentation/ABI/testing/sysfs-kernel-livepatch for more details.
  316. 7. Limitations
  317. ==============
  318. The current Livepatch implementation has several limitations:
  319. - Only functions that can be traced could be patched.
  320. Livepatch is based on the dynamic ftrace. In particular, functions
  321. implementing ftrace or the livepatch ftrace handler could not be
  322. patched. Otherwise, the code would end up in an infinite loop. A
  323. potential mistake is prevented by marking the problematic functions
  324. by "notrace".
  325. - Livepatch works reliably only when the dynamic ftrace is located at
  326. the very beginning of the function.
  327. The function need to be redirected before the stack or the function
  328. parameters are modified in any way. For example, livepatch requires
  329. using -fentry gcc compiler option on x86_64.
  330. One exception is the PPC port. It uses relative addressing and TOC.
  331. Each function has to handle TOC and save LR before it could call
  332. the ftrace handler. This operation has to be reverted on return.
  333. Fortunately, the generic ftrace code has the same problem and all
  334. this is handled on the ftrace level.
  335. - Kretprobes using the ftrace framework conflict with the patched
  336. functions.
  337. Both kretprobes and livepatches use a ftrace handler that modifies
  338. the return address. The first user wins. Either the probe or the patch
  339. is rejected when the handler is already in use by the other.
  340. - Kprobes in the original function are ignored when the code is
  341. redirected to the new implementation.
  342. There is a work in progress to add warnings about this situation.