mirror of
https://github.com/torvalds/linux.git
synced 2024-11-10 06:01:57 +00:00
72ffee678f
The KTSAN doc has moved to https://github.com/google/kernel-sanitizers/blob/master/KTSAN.md. Update the url in kcsan.rst accordingly. Signed-off-by: Haoyang Liu <tttturtleruss@hust.edu.cn> Reviewed-by: Dongliang Mu <dzm91@hust.edu.cn> Acked-by: Marco Elver <elver@google.com> Signed-off-by: Jonathan Corbet <corbet@lwn.net> Link: https://lore.kernel.org/r/20240725174632.23803-1-tttturtleruss@hust.edu.cn
378 lines
17 KiB
ReStructuredText
378 lines
17 KiB
ReStructuredText
.. SPDX-License-Identifier: GPL-2.0
|
|
.. Copyright (C) 2019, Google LLC.
|
|
|
|
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 by both GCC and Clang. With GCC we require version 11 or
|
|
later, and with Clang also require version 11 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 test_kernel_read / test_kernel_write
|
|
|
|
write to 0xffffffffc009a628 of 8 bytes by task 487 on cpu 0:
|
|
test_kernel_write+0x1d/0x30
|
|
access_thread+0x89/0xd0
|
|
kthread+0x23e/0x260
|
|
ret_from_fork+0x22/0x30
|
|
|
|
read to 0xffffffffc009a628 of 8 bytes by task 488 on cpu 6:
|
|
test_kernel_read+0x10/0x20
|
|
access_thread+0x89/0xd0
|
|
kthread+0x23e/0x260
|
|
ret_from_fork+0x22/0x30
|
|
|
|
value changed: 0x00000000000009a6 -> 0x00000000000009b2
|
|
|
|
Reported by Kernel Concurrency Sanitizer on:
|
|
CPU: 6 PID: 488 Comm: access_thread Not tainted 5.12.0-rc2+ #1
|
|
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-2 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. If KCSAN also observed a value change, the observed
|
|
old value and new value are shown on the "value changed" line respectively.
|
|
|
|
The other less common type of data race report looks like this::
|
|
|
|
==================================================================
|
|
BUG: KCSAN: data-race in test_kernel_rmw_array+0x71/0xd0
|
|
|
|
race at unknown origin, with read to 0xffffffffc009bdb0 of 8 bytes by task 515 on cpu 2:
|
|
test_kernel_rmw_array+0x71/0xd0
|
|
access_thread+0x89/0xd0
|
|
kthread+0x23e/0x260
|
|
ret_from_fork+0x22/0x30
|
|
|
|
value changed: 0x0000000000002328 -> 0x0000000000002329
|
|
|
|
Reported by Kernel Concurrency Sanitizer on:
|
|
CPU: 2 PID: 515 Comm: access_thread Not tainted 5.12.0-rc2+ #1
|
|
Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-2 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 reports always show a "value changed"
|
|
line. A common reason for reports of this type are missing instrumentation in
|
|
the racing thread, but could also occur due to e.g. DMA accesses. Such reports
|
|
are shown only if ``CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=y``, which is
|
|
enabled 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. Please see
|
|
`"Marking Shared-Memory Accesses" in the LKMM`_ for more information.
|
|
|
|
* Similar to ``data_race(...)``, the type qualifier ``__data_racy`` can be used
|
|
to document that all data races due to accesses to a variable are intended
|
|
and should be ignored by KCSAN::
|
|
|
|
struct foo {
|
|
...
|
|
int __data_racy stats_counter;
|
|
...
|
|
};
|
|
|
|
* 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.
|
|
|
|
* 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
|
|
|
|
.. _"Marking Shared-Memory Accesses" in the LKMM: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/memory-model/Documentation/access-marking.txt
|
|
|
|
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.
|
|
|
|
* ``CONFIG_KCSAN_PERMISSIVE``: Enable additional permissive rules to ignore
|
|
certain classes of common data races. Unlike the above, the rules are more
|
|
complex involving value-change patterns, access type, and address. This
|
|
option depends on ``CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY=y``. For details
|
|
please see the ``kernel/kcsan/permissive.h``. Testers and maintainers that
|
|
only focus on reports from specific subsystems and not the whole kernel are
|
|
recommended to disable this option.
|
|
|
|
To use the strictest possible rules, select ``CONFIG_KCSAN_STRICT=y``, which
|
|
configures KCSAN to follow the Linux-kernel memory consistency model (LKMM) as
|
|
closely as possible.
|
|
|
|
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.), and a subset of ordering guarantees implied by memory
|
|
barriers. With ``CONFIG_KCSAN_WEAK_MEMORY=y``, KCSAN models load or store
|
|
buffering, and can detect missing ``smp_mb()``, ``smp_wmb()``, ``smp_rmb()``,
|
|
``smp_store_release()``, and all ``atomic_*`` operations with equivalent
|
|
implied barriers.
|
|
|
|
Note, KCSAN will not report all data races due to missing memory ordering,
|
|
specifically where a memory barrier would be required to prohibit subsequent
|
|
memory operation from reordering before the barrier. Developers should
|
|
therefore carefully consider the required memory ordering requirements that
|
|
remain unchecked.
|
|
|
|
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_WRITER_SCOPED
|
|
ASSERT_EXCLUSIVE_ACCESS ASSERT_EXCLUSIVE_ACCESS_SCOPED
|
|
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.
|
|
|
|
Modeling Weak Memory
|
|
~~~~~~~~~~~~~~~~~~~~
|
|
|
|
KCSAN's approach to detecting data races due to missing memory barriers is
|
|
based on modeling access reordering (with ``CONFIG_KCSAN_WEAK_MEMORY=y``).
|
|
Each plain memory access for which a watchpoint is set up, is also selected for
|
|
simulated reordering within the scope of its function (at most 1 in-flight
|
|
access).
|
|
|
|
Once an access has been selected for reordering, it is checked along every
|
|
other access until the end of the function scope. If an appropriate memory
|
|
barrier is encountered, the access will no longer be considered for simulated
|
|
reordering.
|
|
|
|
When the result of a memory operation should be ordered by a barrier, KCSAN can
|
|
then detect data races where the conflict only occurs as a result of a missing
|
|
barrier. Consider the example::
|
|
|
|
int x, flag;
|
|
void T1(void)
|
|
{
|
|
x = 1; // data race!
|
|
WRITE_ONCE(flag, 1); // correct: smp_store_release(&flag, 1)
|
|
}
|
|
void T2(void)
|
|
{
|
|
while (!READ_ONCE(flag)); // correct: smp_load_acquire(&flag)
|
|
... = x; // data race!
|
|
}
|
|
|
|
When weak memory modeling is enabled, KCSAN can consider ``x`` in ``T1`` for
|
|
simulated reordering. After the write of ``flag``, ``x`` is again checked for
|
|
concurrent accesses: because ``T2`` is able to proceed after the write of
|
|
``flag``, a data race is detected. With the correct barriers in place, ``x``
|
|
would not be considered for reordering after the proper release of ``flag``,
|
|
and no data race would be detected.
|
|
|
|
Deliberate trade-offs in complexity but also practical limitations mean only a
|
|
subset of data races due to missing memory barriers can be detected. With
|
|
currently available compiler support, the implementation is limited to modeling
|
|
the effects of "buffering" (delaying accesses), since the runtime cannot
|
|
"prefetch" accesses. Also recall that watchpoints are only set up for plain
|
|
accesses, and the only access type for which KCSAN simulates reordering. This
|
|
means reordering of marked accesses is not modeled.
|
|
|
|
A consequence of the above is that acquire operations do not require barrier
|
|
instrumentation (no prefetching). Furthermore, marked accesses introducing
|
|
address or control dependencies do not require special handling (the marked
|
|
access cannot be reordered, later dependent accesses cannot be prefetched).
|
|
|
|
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 aware of only a subset of LKMM 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/kernel-sanitizers/blob/master/KTSAN.md>`_.
|
|
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.
|