linux/Documentation/dev-tools/kcsan.rst

The Kernel Concurrency Sanitizer (KCSAN)
========================================

The Kernel Concurrency Sanitizer (KCSAN) is a dynamic race detector, which
relies on compile-time instrumentation, and uses a watchpoint-based sampling
approach to detect races. KCSAN's primary purpose is to detect `data races`_.

Usage
-----

KCSAN is supported in both GCC and Clang. With GCC it requires version 7.3.0 or
later. With Clang it requires version 7.0.0 or later.

To enable KCSAN configure the kernel with::

    CONFIG_KCSAN = y

KCSAN provides several other configuration options to customize behaviour (see
the respective help text in ``lib/Kconfig.kcsan`` for more info).

Error reports
~~~~~~~~~~~~~

A typical data race report looks like this::

    ==================================================================
    BUG: KCSAN: data-race in generic_permission / kernfs_refresh_inode

    write to 0xffff8fee4c40700c of 4 bytes by task 175 on cpu 4:
     kernfs_refresh_inode+0x70/0x170
     kernfs_iop_permission+0x4f/0x90
     inode_permission+0x190/0x200
     link_path_walk.part.0+0x503/0x8e0
     path_lookupat.isra.0+0x69/0x4d0
     filename_lookup+0x136/0x280
     user_path_at_empty+0x47/0x60
     vfs_statx+0x9b/0x130
     __do_sys_newlstat+0x50/0xb0
     __x64_sys_newlstat+0x37/0x50
     do_syscall_64+0x85/0x260
     entry_SYSCALL_64_after_hwframe+0x44/0xa9

    read to 0xffff8fee4c40700c of 4 bytes by task 166 on cpu 6:
     generic_permission+0x5b/0x2a0
     kernfs_iop_permission+0x66/0x90
     inode_permission+0x190/0x200
     link_path_walk.part.0+0x503/0x8e0
     path_lookupat.isra.0+0x69/0x4d0
     filename_lookup+0x136/0x280
     user_path_at_empty+0x47/0x60
     do_faccessat+0x11a/0x390
     __x64_sys_access+0x3c/0x50
     do_syscall_64+0x85/0x260
     entry_SYSCALL_64_after_hwframe+0x44/0xa9

    Reported by Kernel Concurrency Sanitizer on:
    CPU: 6 PID: 166 Comm: systemd-journal Not tainted 5.3.0-rc7+ #1
    Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
    ==================================================================

The header of the report provides a short summary of the functions involved in
the race. It is followed by the access types and stack traces of the 2 threads
involved in the data race.

The other less common type of data race report looks like this::

    ==================================================================
    BUG: KCSAN: data-race in e1000_clean_rx_irq+0x551/0xb10

    race at unknown origin, with read to 0xffff933db8a2ae6c of 1 bytes by interrupt on cpu 0:
     e1000_clean_rx_irq+0x551/0xb10
     e1000_clean+0x533/0xda0
     net_rx_action+0x329/0x900
     __do_softirq+0xdb/0x2db
     irq_exit+0x9b/0xa0
     do_IRQ+0x9c/0xf0
     ret_from_intr+0x0/0x18
     default_idle+0x3f/0x220
     arch_cpu_idle+0x21/0x30
     do_idle+0x1df/0x230
     cpu_startup_entry+0x14/0x20
     rest_init+0xc5/0xcb
     arch_call_rest_init+0x13/0x2b
     start_kernel+0x6db/0x700

    Reported by Kernel Concurrency Sanitizer on:
    CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.3.0-rc7+ #2
    Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
    ==================================================================

This report is generated where it was not possible to determine the other
racing thread, but a race was inferred due to the data value of the watched
memory location having changed. These can occur either due to missing
instrumentation or e.g. DMA accesses. These reports will only be generated if
``CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=y`` (selected by default).

Selective analysis
~~~~~~~~~~~~~~~~~~

It may be desirable to disable data race detection for specific accesses,
functions, compilation units, or entire subsystems.  For static blacklisting,
the below options are available:

* KCSAN understands the ``data_race(expr)`` annotation, which tells KCSAN that
  any data races due to accesses in ``expr`` should be ignored and resulting
  behaviour when encountering a data race is deemed safe.

* Disabling data race detection for entire functions can be accomplished by
  using the function attribute ``__no_kcsan``::

    __no_kcsan
    void foo(void) {
        ...

  To dynamically limit for which functions to generate reports, see the
  `DebugFS interface`_ blacklist/whitelist feature.

  For ``__always_inline`` functions, replace ``__always_inline`` with
  ``__no_kcsan_or_inline`` (which implies ``__always_inline``)::

    static __no_kcsan_or_inline void foo(void) {
        ...

  Note: Older compiler versions (GCC < 9) also do not always honor the
  ``__no_kcsan`` attribute on regular ``inline`` functions. If false positives
  with these compilers cannot be tolerated, for small functions where
  ``__always_inline`` would be appropriate, ``__no_kcsan_or_inline`` should be
  preferred instead.

* To disable data race detection for a particular compilation unit, add to the
  ``Makefile``::

    KCSAN_SANITIZE_file.o := n

* To disable data race detection for all compilation units listed in a
  ``Makefile``, add to the respective ``Makefile``::

    KCSAN_SANITIZE := n

Furthermore, it is possible to tell KCSAN to show or hide entire classes of
data races, depending on preferences. These can be changed via the following
Kconfig options:

* ``CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY``: If enabled and a conflicting write
  is observed via a watchpoint, but the data value of the memory location was
  observed to remain unchanged, do not report the data race.

* ``CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC``: Assume that plain aligned writes
  up to word size are atomic by default. Assumes that such writes are not
  subject to unsafe compiler optimizations resulting in data races. The option
  causes KCSAN to not report data races due to conflicts where the only plain
  accesses are aligned writes up to word size.

DebugFS interface
~~~~~~~~~~~~~~~~~

The file ``/sys/kernel/debug/kcsan`` provides the following interface:

* Reading ``/sys/kernel/debug/kcsan`` returns various runtime statistics.

* Writing ``on`` or ``off`` to ``/sys/kernel/debug/kcsan`` allows turning KCSAN
  on or off, respectively.

* Writing ``!some_func_name`` to ``/sys/kernel/debug/kcsan`` adds
  ``some_func_name`` to the report filter list, which (by default) blacklists
  reporting data races where either one of the top stackframes are a function
  in the list.

* Writing either ``blacklist`` or ``whitelist`` to ``/sys/kernel/debug/kcsan``
  changes the report filtering behaviour. For example, the blacklist feature
  can be used to silence frequently occurring data races; the whitelist feature
  can help with reproduction and testing of fixes.

Tuning performance
~~~~~~~~~~~~~~~~~~

Core parameters that affect KCSAN's overall performance and bug detection
ability are exposed as kernel command-line arguments whose defaults can also be
changed via the corresponding Kconfig options.

* ``kcsan.skip_watch`` (``CONFIG_KCSAN_SKIP_WATCH``): Number of per-CPU memory
  operations to skip, before another watchpoint is set up. Setting up
  watchpoints more frequently will result in the likelihood of races to be
  observed to increase. This parameter has the most significant impact on
  overall system performance and race detection ability.

* ``kcsan.udelay_task`` (``CONFIG_KCSAN_UDELAY_TASK``): For tasks, the
  microsecond delay to stall execution after a watchpoint has been set up.
  Larger values result in the window in which we may observe a race to
  increase.

* ``kcsan.udelay_interrupt`` (``CONFIG_KCSAN_UDELAY_INTERRUPT``): For
  interrupts, the microsecond delay to stall execution after a watchpoint has
  been set up. Interrupts have tighter latency requirements, and their delay
  should generally be smaller than the one chosen for tasks.

They may be tweaked at runtime via ``/sys/module/kcsan/parameters/``.

Data Races
----------

In an execution, two memory accesses form a *data race* if they *conflict*,
they happen concurrently in different threads, and at least one of them is a
*plain access*; they *conflict* if both access the same memory location, and at
least one is a write. For a more thorough discussion and definition, see `"Plain
Accesses and Data Races" in the LKMM`_.

.. _"Plain Accesses and Data Races" in the LKMM: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/memory-model/Documentation/explanation.txt#n1922

Relationship with the Linux-Kernel Memory Consistency Model (LKMM)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

The LKMM defines the propagation and ordering rules of various memory
operations, which gives developers the ability to reason about concurrent code.
Ultimately this allows to determine the possible executions of concurrent code,
and if that code is free from data races.

KCSAN is aware of *marked atomic operations* (``READ_ONCE``, ``WRITE_ONCE``,
``atomic_*``, etc.), but is oblivious of any ordering guarantees and simply
assumes that memory barriers are placed correctly. In other words, KCSAN
assumes that as long as a plain access is not observed to race with another
conflicting access, memory operations are correctly ordered.

This means that KCSAN will not report *potential* data races due to missing
memory ordering. Developers should therefore carefully consider the required
memory ordering requirements that remain unchecked. If, however, missing
memory ordering (that is observable with a particular compiler and
architecture) leads to an observable data race (e.g. entering a critical
section erroneously), KCSAN would report the resulting data race.

Race Detection Beyond Data Races
--------------------------------

For code with complex concurrency design, race-condition bugs may not always
manifest as data races. Race conditions occur if concurrently executing
operations result in unexpected system behaviour. On the other hand, data races
are defined at the C-language level. The following macros can be used to check
properties of concurrent code where bugs would not manifest as data races.

.. kernel-doc:: include/linux/kcsan-checks.h
    :functions: ASSERT_EXCLUSIVE_WRITER ASSERT_EXCLUSIVE_ACCESS
                ASSERT_EXCLUSIVE_BITS

Implementation Details
----------------------

KCSAN relies on observing that two accesses happen concurrently. Crucially, we
want to (a) increase the chances of observing races (especially for races that
manifest rarely), and (b) be able to actually observe them. We can accomplish
(a) by injecting various delays, and (b) by using address watchpoints (or
breakpoints).

If we deliberately stall a memory access, while we have a watchpoint for its
address set up, and then observe the watchpoint to fire, two accesses to the
same address just raced. Using hardware watchpoints, this is the approach taken
in `DataCollider
<http://usenix.org/legacy/events/osdi10/tech/full_papers/Erickson.pdf>`_.
Unlike DataCollider, KCSAN does not use hardware watchpoints, but instead
relies on compiler instrumentation and "soft watchpoints".

In KCSAN, watchpoints are implemented using an efficient encoding that stores
access type, size, and address in a long; the benefits of using "soft
watchpoints" are portability and greater flexibility. KCSAN then relies on the
compiler instrumenting plain accesses. For each instrumented plain access:

1. Check if a matching watchpoint exists; if yes, and at least one access is a
   write, then we encountered a racing access.

2. Periodically, if no matching watchpoint exists, set up a watchpoint and
   stall for a small randomized delay.

3. Also check the data value before the delay, and re-check the data value
   after delay; if the values mismatch, we infer a race of unknown origin.

To detect data races between plain and marked accesses, KCSAN also annotates
marked accesses, but only to check if a watchpoint exists; i.e. KCSAN never
sets up a watchpoint on marked accesses. By never setting up watchpoints for
marked operations, if all accesses to a variable that is accessed concurrently
are properly marked, KCSAN will never trigger a watchpoint and therefore never
report the accesses.

Key Properties
~~~~~~~~~~~~~~

1. **Memory Overhead:**  The overall memory overhead is only a few MiB
   depending on configuration. The current implementation uses a small array of
   longs to encode watchpoint information, which is negligible.

2. **Performance Overhead:** KCSAN's runtime aims to be minimal, using an
   efficient watchpoint encoding that does not require acquiring any shared
   locks in the fast-path. For kernel boot on a system with 8 CPUs:

   - 5.0x slow-down with the default KCSAN config;
   - 2.8x slow-down from runtime fast-path overhead only (set very large
     ``KCSAN_SKIP_WATCH`` and unset ``KCSAN_SKIP_WATCH_RANDOMIZE``).

3. **Annotation Overheads:** Minimal annotations are required outside the KCSAN
   runtime. As a result, maintenance overheads are minimal as the kernel
   evolves.

4. **Detects Racy Writes from Devices:** Due to checking data values upon
   setting up watchpoints, racy writes from devices can also be detected.

5. **Memory Ordering:** KCSAN is *not* explicitly aware of the LKMM's ordering
   rules; this may result in missed data races (false negatives).

6. **Analysis Accuracy:** For observed executions, due to using a sampling
   strategy, the analysis is *unsound* (false negatives possible), but aims to
   be complete (no false positives).

Alternatives Considered
-----------------------

An alternative data race detection approach for the kernel can be found in the
`Kernel Thread Sanitizer (KTSAN) <https://github.com/google/ktsan/wiki>`_.
KTSAN is a happens-before data race detector, which explicitly establishes the
happens-before order between memory operations, which can then be used to
determine data races as defined in `Data Races`_.

To build a correct happens-before relation, KTSAN must be aware of all ordering
rules of the LKMM and synchronization primitives. Unfortunately, any omission
leads to large numbers of false positives, which is especially detrimental in
the context of the kernel which includes numerous custom synchronization
mechanisms. To track the happens-before relation, KTSAN's implementation
requires metadata for each memory location (shadow memory), which for each page
corresponds to 4 pages of shadow memory, and can translate into overhead of
tens of GiB on a large system.