=========
  Livepatch
  =========
  
  This document outlines basic information about kernel livepatching.
  
  Table of Contents:
  
  1. Motivation
  2. Kprobes, Ftrace, Livepatching
  3. Consistency model
  4. Livepatch module
     4.1. New functions
     4.2. Metadata
     4.3. Livepatch module handling
  5. Livepatch life-cycle
     5.1. Registration
     5.2. Enabling
     5.3. Disabling
     5.4. Unregistration
  6. Sysfs
  7. Limitations
  
  
  1. Motivation
  =============
  
  There are many situations where users are reluctant to reboot a system. It may
  be because their system is performing complex scientific computations or under
  heavy load during peak usage. In addition to keeping systems up and running,
  users want to also have a stable and secure system. Livepatching gives users
  both by allowing for function calls to be redirected; thus, fixing critical
  functions without a system reboot.
  
  
  2. Kprobes, Ftrace, Livepatching
  ================================
  
  There are multiple mechanisms in the Linux kernel that are directly related
  to redirection of code execution; namely: kernel probes, function tracing,
  and livepatching:
  
    + The kernel probes are the most generic. The code can be redirected by
      putting a breakpoint instruction instead of any instruction.
  
    + The function tracer calls the code from a predefined location that is
      close to the function entry point. This location is generated by the
      compiler using the '-pg' gcc option.
  
    + Livepatching typically needs to redirect the code at the very beginning
      of the function entry before the function parameters or the stack
      are in any way modified.
  
  All three approaches need to modify the existing code at runtime. Therefore
  they need to be aware of each other and not step over each other's toes.
  Most of these problems are solved by using the dynamic ftrace framework as
  a base. A Kprobe is registered as a ftrace handler when the function entry
  is probed, see CONFIG_KPROBES_ON_FTRACE. Also an alternative function from
  a live patch is called with the help of a custom ftrace handler. But there are
  some limitations, see below.
  
  
  3. Consistency model
  ====================
  
  Functions are there for a reason. They take some input parameters, get or
  release locks, read, process, and even write some data in a defined way,
  have return values. In other words, each function has a defined semantic.
  
  Many fixes do not change the semantic of the modified functions. For
  example, they add a NULL pointer or a boundary check, fix a race by adding
  a missing memory barrier, or add some locking around a critical section.
  Most of these changes are self contained and the function presents itself
  the same way to the rest of the system. In this case, the functions might

  be updated independently one by one.  (This can be done by setting the
  'immediate' flag in the klp_patch struct.)

  
  But there are more complex fixes. For example, a patch might change
  ordering of locking in multiple functions at the same time. Or a patch
  might exchange meaning of some temporary structures and update
  all the relevant functions. In this case, the affected unit
  (thread, whole kernel) need to start using all new versions of
  the functions at the same time. Also the switch must happen only
  when it is safe to do so, e.g. when the affected locks are released
  or no data are stored in the modified structures at the moment.
  
  The theory about how to apply functions a safe way is rather complex.
  The aim is to define a so-called consistency model. It attempts to define
  conditions when the new implementation could be used so that the system

  stays consistent.
  
  Livepatch has a consistency model which is a hybrid of kGraft and
  kpatch:  it uses kGraft's per-task consistency and syscall barrier
  switching combined with kpatch's stack trace switching.  There are also
  a number of fallback options which make it quite flexible.
  
  Patches are applied on a per-task basis, when the task is deemed safe to
  switch over.  When a patch is enabled, livepatch enters into a
  transition state where tasks are converging to the patched state.
  Usually this transition state can complete in a few seconds.  The same
  sequence occurs when a patch is disabled, except the tasks converge from
  the patched state to the unpatched state.
  
  An interrupt handler inherits the patched state of the task it
  interrupts.  The same is true for forked tasks: the child inherits the
  patched state of the parent.
  
  Livepatch uses several complementary approaches to determine when it's
  safe to patch tasks:
  
  1. The first and most effective approach is stack checking of sleeping
     tasks.  If no affected functions are on the stack of a given task,
     the task is patched.  In most cases this will patch most or all of
     the tasks on the first try.  Otherwise it'll keep trying
     periodically.  This option is only available if the architecture has
     reliable stacks (HAVE_RELIABLE_STACKTRACE).
  
  2. The second approach, if needed, is kernel exit switching.  A
     task is switched when it returns to user space from a system call, a
     user space IRQ, or a signal.  It's useful in the following cases:
  
     a) Patching I/O-bound user tasks which are sleeping on an affected
        function.  In this case you have to send SIGSTOP and SIGCONT to
        force it to exit the kernel and be patched.
     b) Patching CPU-bound user tasks.  If the task is highly CPU-bound
        then it will get patched the next time it gets interrupted by an
        IRQ.
     c) In the future it could be useful for applying patches for
        architectures which don't yet have HAVE_RELIABLE_STACKTRACE.  In
        this case you would have to signal most of the tasks on the
        system.  However this isn't supported yet because there's
        currently no way to patch kthreads without
        HAVE_RELIABLE_STACKTRACE.
  
  3. For idle "swapper" tasks, since they don't ever exit the kernel, they
     instead have a klp_update_patch_state() call in the idle loop which
     allows them to be patched before the CPU enters the idle state.
  
     (Note there's not yet such an approach for kthreads.)
  
  All the above approaches may be skipped by setting the 'immediate' flag
  in the 'klp_patch' struct, which will disable per-task consistency and
  patch all tasks immediately.  This can be useful if the patch doesn't
  change any function or data semantics.  Note that, even with this flag
  set, it's possible that some tasks may still be running with an old
  version of the function, until that function returns.
  
  There's also an 'immediate' flag in the 'klp_func' struct which allows
  you to specify that certain functions in the patch can be applied
  without per-task consistency.  This might be useful if you want to patch
  a common function like schedule(), and the function change doesn't need
  consistency but the rest of the patch does.
  
  For architectures which don't have HAVE_RELIABLE_STACKTRACE, the user
  must set patch->immediate which causes all tasks to be patched
  immediately.  This option should be used with care, only when the patch
  doesn't change any function or data semantics.
  
  In the future, architectures which don't have HAVE_RELIABLE_STACKTRACE
  may be allowed to use per-task consistency if we can come up with
  another way to patch kthreads.
  
  The /sys/kernel/livepatch/<patch>/transition file shows whether a patch
  is in transition.  Only a single patch (the topmost patch on the stack)
  can be in transition at a given time.  A patch can remain in transition
  indefinitely, if any of the tasks are stuck in the initial patch state.
  
  A transition can be reversed and effectively canceled by writing the
  opposite value to the /sys/kernel/livepatch/<patch>/enabled file while
  the transition is in progress.  Then all the tasks will attempt to
  converge back to the original patch state.
  
  There's also a /proc/<pid>/patch_state file which can be used to
  determine which tasks are blocking completion of a patching operation.
  If a patch is in transition, this file shows 0 to indicate the task is
  unpatched and 1 to indicate it's patched.  Otherwise, if no patch is in
  transition, it shows -1.  Any tasks which are blocking the transition
  can be signaled with SIGSTOP and SIGCONT to force them to change their
  patched state.
  
  
  3.1 Adding consistency model support to new architectures
  ---------------------------------------------------------
  
  For adding consistency model support to new architectures, there are a
  few options:
  
  1) Add CONFIG_HAVE_RELIABLE_STACKTRACE.  This means porting objtool, and
     for non-DWARF unwinders, also making sure there's a way for the stack
     tracing code to detect interrupts on the stack.
  
  2) Alternatively, ensure that every kthread has a call to
     klp_update_patch_state() in a safe location.  Kthreads are typically
     in an infinite loop which does some action repeatedly.  The safe
     location to switch the kthread's patch state would be at a designated
     point in the loop where there are no locks taken and all data
     structures are in a well-defined state.
  
     The location is clear when using workqueues or the kthread worker
     API.  These kthreads process independent actions in a generic loop.
  
     It's much more complicated with kthreads which have a custom loop.
     There the safe location must be carefully selected on a case-by-case
     basis.
  
     In that case, arches without HAVE_RELIABLE_STACKTRACE would still be
     able to use the non-stack-checking parts of the consistency model:
  
     a) patching user tasks when they cross the kernel/user space
        boundary; and
  
     b) patching kthreads and idle tasks at their designated patch points.
  
     This option isn't as good as option 1 because it requires signaling
     user tasks and waking kthreads to patch them.  But it could still be
     a good backup option for those architectures which don't have
     reliable stack traces yet.
  
  In the meantime, patches for such architectures can bypass the
  consistency model by setting klp_patch.immediate to true.  This option
  is perfectly fine for patches which don't change the semantics of the
  patched functions.  In practice, this is usable for ~90% of security
  fixes.  Use of this option also means the patch can't be unloaded after
  it has been disabled.

  
  
  4. Livepatch module
  ===================
  
  Livepatches are distributed using kernel modules, see
  samples/livepatch/livepatch-sample.c.
  
  The module includes a new implementation of functions that we want
  to replace. In addition, it defines some structures describing the
  relation between the original and the new implementation. Then there
  is code that makes the kernel start using the new code when the livepatch
  module is loaded. Also there is code that cleans up before the
  livepatch module is removed. All this is explained in more details in
  the next sections.
  
  
  4.1. New functions
  ------------------
  
  New versions of functions are typically just copied from the original
  sources. A good practice is to add a prefix to the names so that they
  can be distinguished from the original ones, e.g. in a backtrace. Also
  they can be declared as static because they are not called directly
  and do not need the global visibility.
  
  The patch contains only functions that are really modified. But they
  might want to access functions or data from the original source file
  that may only be locally accessible. This can be solved by a special
  relocation section in the generated livepatch module, see
  Documentation/livepatch/module-elf-format.txt for more details.
  
  
  4.2. Metadata

  -------------

  
  The patch is described by several structures that split the information
  into three levels:
  
    + struct klp_func is defined for each patched function. It describes
      the relation between the original and the new implementation of a
      particular function.
  
      The structure includes the name, as a string, of the original function.
      The function address is found via kallsyms at runtime.
  
      Then it includes the address of the new function. It is defined
      directly by assigning the function pointer. Note that the new
      function is typically defined in the same source file.
  
      As an optional parameter, the symbol position in the kallsyms database can
      be used to disambiguate functions of the same name. This is not the
      absolute position in the database, but rather the order it has been found
      only for a particular object ( vmlinux or a kernel module ). Note that
      kallsyms allows for searching symbols according to the object name.

      There's also an 'immediate' flag which, when set, patches the
      function immediately, bypassing the consistency model safety checks.

    + struct klp_object defines an array of patched functions (struct
      klp_func) in the same object. Where the object is either vmlinux
      (NULL) or a module name.
  
      The structure helps to group and handle functions for each object
      together. Note that patched modules might be loaded later than
      the patch itself and the relevant functions might be patched
      only when they are available.
  
  
    + struct klp_patch defines an array of patched objects (struct
      klp_object).
  
      This structure handles all patched functions consistently and eventually,
      synchronously. The whole patch is applied only when all patched
      symbols are found. The only exception are symbols from objects

      (kernel modules) that have not been loaded yet.
  
      Setting the 'immediate' flag applies the patch to all tasks
      immediately, bypassing the consistency model safety checks.
  
      For more details on how the patch is applied on a per-task basis,
      see the "Consistency model" section.

  
  
  4.3. Livepatch module handling
  ------------------------------
  
  The usual behavior is that the new functions will get used when
  the livepatch module is loaded. For this, the module init() function
  has to register the patch (struct klp_patch) and enable it. See the
  section "Livepatch life-cycle" below for more details about these
  two operations.
  
  Module removal is only safe when there are no users of the underlying

  functions. The immediate consistency model is not able to detect this. The
  code just redirects the functions at the very beginning and it does not
  check if the functions are in use. In other words, it knows when the
  functions get called but it does not know when the functions return.
  Therefore it cannot be decided when the livepatch module can be safely
  removed. This is solved by a hybrid consistency model. When the system is
  transitioned to a new patch state (patched/unpatched) it is guaranteed that
  no task sleeps or runs in the old code.

  
  5. Livepatch life-cycle
  =======================
  
  Livepatching defines four basic operations that define the life cycle of each
  live patch: registration, enabling, disabling and unregistration.  There are
  several reasons why it is done this way.
  
  First, the patch is applied only when all patched symbols for already
  loaded objects are found. The error handling is much easier if this
  check is done before particular functions get redirected.
  
  Second, the immediate consistency model does not guarantee that anyone is not
  sleeping in the new code after the patch is reverted. This means that the new
  code needs to stay around "forever". If the code is there, one could apply it
  again. Therefore it makes sense to separate the operations that might be done
  once and those that need to be repeated when the patch is enabled (applied)
  again.
  
  Third, it might take some time until the entire system is migrated
  when a more complex consistency model is used. The patch revert might
  block the livepatch module removal for too long. Therefore it is useful
  to revert the patch using a separate operation that might be called
  explicitly. But it does not make sense to remove all information
  until the livepatch module is really removed.
  
  
  5.1. Registration
  -----------------
  
  Each patch first has to be registered using klp_register_patch(). This makes
  the patch known to the livepatch framework. Also it does some preliminary
  computing and checks.
  
  In particular, the patch is added into the list of known patches. The
  addresses of the patched functions are found according to their names.
  The special relocations, mentioned in the section "New functions", are
  applied. The relevant entries are created under
  /sys/kernel/livepatch/<name>. The patch is rejected when any operation
  fails.
  
  
  5.2. Enabling
  -------------
  
  Registered patches might be enabled either by calling klp_enable_patch() or
  by writing '1' to /sys/kernel/livepatch/<name>/enabled. The system will
  start using the new implementation of the patched functions at this stage.

  When a patch is enabled, livepatch enters into a transition state where
  tasks are converging to the patched state.  This is indicated by a value
  of '1' in /sys/kernel/livepatch/<name>/transition.  Once all tasks have
  been patched, the 'transition' value changes to '0'.  For more
  information about this process, see the "Consistency model" section.
  
  If an original function is patched for the first time, a function
  specific struct klp_ops is created and an universal ftrace handler is
  registered.

  
  Functions might be patched multiple times. The ftrace handler is registered
  only once for the given function. Further patches just add an entry to the
  list (see field `func_stack`) of the struct klp_ops. The last added
  entry is chosen by the ftrace handler and becomes the active function
  replacement.
  
  Note that the patches might be enabled in a different order than they were
  registered.
  
  
  5.3. Disabling
  --------------
  
  Enabled patches might get disabled either by calling klp_disable_patch() or
  by writing '0' to /sys/kernel/livepatch/<name>/enabled. At this stage
  either the code from the previously enabled patch or even the original
  code gets used.

  When a patch is disabled, livepatch enters into a transition state where
  tasks are converging to the unpatched state.  This is indicated by a
  value of '1' in /sys/kernel/livepatch/<name>/transition.  Once all tasks
  have been unpatched, the 'transition' value changes to '0'.  For more
  information about this process, see the "Consistency model" section.

  Here all the functions (struct klp_func) associated with the to-be-disabled
  patch are removed from the corresponding struct klp_ops. The ftrace handler
  is unregistered and the struct klp_ops is freed when the func_stack list
  becomes empty.
  
  Patches must be disabled in exactly the reverse order in which they were
  enabled. It makes the problem and the implementation much easier.
  
  
  5.4. Unregistration
  -------------------
  
  Disabled patches might be unregistered by calling klp_unregister_patch().
  This can be done only when the patch is disabled and the code is no longer
  used. It must be called before the livepatch module gets unloaded.
  
  At this stage, all the relevant sys-fs entries are removed and the patch
  is removed from the list of known patches.
  
  
  6. Sysfs
  ========
  
  Information about the registered patches can be found under
  /sys/kernel/livepatch. The patches could be enabled and disabled
  by writing there.
  
  See Documentation/ABI/testing/sysfs-kernel-livepatch for more details.
  
  
  7. Limitations
  ==============
  
  The current Livepatch implementation has several limitations:
  
  
    + The patch must not change the semantic of the patched functions.
  
      The current implementation guarantees only that either the old
      or the new function is called. The functions are patched one
      by one. It means that the patch must _not_ change the semantic
      of the function.
  
  
    + Data structures can not be patched.
  
      There is no support to version data structures or anyhow migrate
      one structure into another. Also the simple consistency model does
      not allow to switch more functions atomically.
  
      Once there is more complex consistency mode, it will be possible to
      use some workarounds. For example, it will be possible to use a hole
      for a new member because the data structure is aligned. Or it will
      be possible to use an existing member for something else.
  
      There are no plans to add more generic support for modified structures
      at the moment.
  
  
    + Only functions that can be traced could be patched.
  
      Livepatch is based on the dynamic ftrace. In particular, functions
      implementing ftrace or the livepatch ftrace handler could not be
      patched. Otherwise, the code would end up in an infinite loop. A
      potential mistake is prevented by marking the problematic functions
      by "notrace".

  
    + Livepatch works reliably only when the dynamic ftrace is located at
      the very beginning of the function.
  
      The function need to be redirected before the stack or the function
      parameters are modified in any way. For example, livepatch requires
      using -fentry gcc compiler option on x86_64.
  
      One exception is the PPC port. It uses relative addressing and TOC.
      Each function has to handle TOC and save LR before it could call
      the ftrace handler. This operation has to be reverted on return.
      Fortunately, the generic ftrace code has the same problem and all

      this is handled on the ftrace level.

  
  
    + Kretprobes using the ftrace framework conflict with the patched
      functions.
  
      Both kretprobes and livepatches use a ftrace handler that modifies
      the return address. The first user wins. Either the probe or the patch
      is rejected when the handler is already in use by the other.
  
  
    + Kprobes in the original function are ignored when the code is
      redirected to the new implementation.
  
      There is a work in progress to add warnings about this situation.