linux-next/Documentation/livepatch/livepatch.txt

=========
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.

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.

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.)

Architectures which don't have HAVE_RELIABLE_STACKTRACE solely rely on
the second approach. It's highly likely that some tasks may still be
running with an old version of the function, until that function
returns. In this case you would have to signal the tasks. This
especially applies to kthreads. They may not be woken up and would need
to be forced. See below for more information.

Unless we can come up with another way to patch kthreads, architectures
without HAVE_RELIABLE_STACKTRACE are not considered fully supported by
the kernel livepatching.

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. This may be harmful to the system though.
/sys/kernel/livepatch/<patch>/signal attribute provides a better alternative.
Writing 1 to the attribute sends a fake signal to all remaining blocking
tasks. No proper signal is actually delivered (there is no data in signal
pending structures). Tasks are interrupted or woken up, and forced to change
their patched state.

Administrator can also affect a transition through
/sys/kernel/livepatch/<patch>/force attribute. Writing 1 there clears
TIF_PATCH_PENDING flag of all tasks and thus forces the tasks to the patched
state. Important note! The force attribute is intended for cases when the
transition gets stuck for a long time because of a blocking task. Administrator
is expected to collect all necessary data (namely stack traces of such blocking
tasks) and request a clearance from a patch distributor to force the transition.
Unauthorized usage may cause harm to the system. It depends on the nature of the
patch, which functions are (un)patched, and which functions the blocking tasks
are sleeping in (/proc/<pid>/stack may help here). Removal (rmmod) of patch
modules is permanently disabled when the force feature is used. It cannot be
guaranteed there is no task sleeping in such module. It implies unbounded
reference count if a patch module is disabled and enabled in a loop.

Moreover, the usage of force may also affect future applications of live
patches and cause even more harm to the system. Administrator should first
consider to simply cancel a transition (see above). If force is used, reboot
should be planned and no more live patches applied.

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.


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.

  + 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.

    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. This is the reason why the force feature permanently disables
the removal. The forced tasks entered the functions but we cannot say
that they returned back.  Therefore it cannot be decided when the
livepatch module can be safely removed. When the system is successfully
transitioned to a new patch state (patched/unpatched) without being
forced 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, it might take some time until the entire system is migrated with
the hybrid consistency model being 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.

/sys/kernel/livepatch/<patch>/signal and /sys/kernel/livepatch/<patch>/force
attributes allow administrator to affect a patching operation.

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.