1 .. SPDX-License-Identifier: GPL-2.0
2 .. Copyright (C) 2019, Google LLC.
4 The Kernel Concurrency Sanitizer (KCSAN)
5 ========================================
7 The Kernel Concurrency Sanitizer (KCSAN) is a dynamic race detector, which
8 relies on compile-time instrumentation, and uses a watchpoint-based sampling
9 approach to detect races. KCSAN's primary purpose is to detect `data races`_.
14 KCSAN is supported by both GCC and Clang. With GCC we require version 11 or
15 later, and with Clang also require version 11 or later.
17 To enable KCSAN configure the kernel with::
21 KCSAN provides several other configuration options to customize behaviour (see
22 the respective help text in ``lib/Kconfig.kcsan`` for more info).
27 A typical data race report looks like this::
29 ==================================================================
30 BUG: KCSAN: data-race in test_kernel_read / test_kernel_write
32 write to 0xffffffffc009a628 of 8 bytes by task 487 on cpu 0:
33 test_kernel_write+0x1d/0x30
34 access_thread+0x89/0xd0
36 ret_from_fork+0x22/0x30
38 read to 0xffffffffc009a628 of 8 bytes by task 488 on cpu 6:
39 test_kernel_read+0x10/0x20
40 access_thread+0x89/0xd0
42 ret_from_fork+0x22/0x30
44 value changed: 0x00000000000009a6 -> 0x00000000000009b2
46 Reported by Kernel Concurrency Sanitizer on:
47 CPU: 6 PID: 488 Comm: access_thread Not tainted 5.12.0-rc2+ #1
48 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-2 04/01/2014
49 ==================================================================
51 The header of the report provides a short summary of the functions involved in
52 the race. It is followed by the access types and stack traces of the 2 threads
53 involved in the data race. If KCSAN also observed a value change, the observed
54 old value and new value are shown on the "value changed" line respectively.
56 The other less common type of data race report looks like this::
58 ==================================================================
59 BUG: KCSAN: data-race in test_kernel_rmw_array+0x71/0xd0
61 race at unknown origin, with read to 0xffffffffc009bdb0 of 8 bytes by task 515 on cpu 2:
62 test_kernel_rmw_array+0x71/0xd0
63 access_thread+0x89/0xd0
65 ret_from_fork+0x22/0x30
67 value changed: 0x0000000000002328 -> 0x0000000000002329
69 Reported by Kernel Concurrency Sanitizer on:
70 CPU: 2 PID: 515 Comm: access_thread Not tainted 5.12.0-rc2+ #1
71 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.14.0-2 04/01/2014
72 ==================================================================
74 This report is generated where it was not possible to determine the other
75 racing thread, but a race was inferred due to the data value of the watched
76 memory location having changed. These reports always show a "value changed"
77 line. A common reason for reports of this type are missing instrumentation in
78 the racing thread, but could also occur due to e.g. DMA accesses. Such reports
79 are shown only if ``CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=y``, which is
85 It may be desirable to disable data race detection for specific accesses,
86 functions, compilation units, or entire subsystems. For static blacklisting,
87 the below options are available:
89 * KCSAN understands the ``data_race(expr)`` annotation, which tells KCSAN that
90 any data races due to accesses in ``expr`` should be ignored and resulting
91 behaviour when encountering a data race is deemed safe. Please see
92 `"Marking Shared-Memory Accesses" in the LKMM`_ for more information.
94 * Disabling data race detection for entire functions can be accomplished by
95 using the function attribute ``__no_kcsan``::
101 To dynamically limit for which functions to generate reports, see the
102 `DebugFS interface`_ blacklist/whitelist feature.
104 * To disable data race detection for a particular compilation unit, add to the
107 KCSAN_SANITIZE_file.o := n
109 * To disable data race detection for all compilation units listed in a
110 ``Makefile``, add to the respective ``Makefile``::
114 .. _"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
116 Furthermore, it is possible to tell KCSAN to show or hide entire classes of
117 data races, depending on preferences. These can be changed via the following
120 * ``CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY``: If enabled and a conflicting write
121 is observed via a watchpoint, but the data value of the memory location was
122 observed to remain unchanged, do not report the data race.
124 * ``CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC``: Assume that plain aligned writes
125 up to word size are atomic by default. Assumes that such writes are not
126 subject to unsafe compiler optimizations resulting in data races. The option
127 causes KCSAN to not report data races due to conflicts where the only plain
128 accesses are aligned writes up to word size.
130 * ``CONFIG_KCSAN_PERMISSIVE``: Enable additional permissive rules to ignore
131 certain classes of common data races. Unlike the above, the rules are more
132 complex involving value-change patterns, access type, and address. This
133 option depends on ``CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY=y``. For details
134 please see the ``kernel/kcsan/permissive.h``. Testers and maintainers that
135 only focus on reports from specific subsystems and not the whole kernel are
136 recommended to disable this option.
138 To use the strictest possible rules, select ``CONFIG_KCSAN_STRICT=y``, which
139 configures KCSAN to follow the Linux-kernel memory consistency model (LKMM) as
145 The file ``/sys/kernel/debug/kcsan`` provides the following interface:
147 * Reading ``/sys/kernel/debug/kcsan`` returns various runtime statistics.
149 * Writing ``on`` or ``off`` to ``/sys/kernel/debug/kcsan`` allows turning KCSAN
150 on or off, respectively.
152 * Writing ``!some_func_name`` to ``/sys/kernel/debug/kcsan`` adds
153 ``some_func_name`` to the report filter list, which (by default) blacklists
154 reporting data races where either one of the top stackframes are a function
157 * Writing either ``blacklist`` or ``whitelist`` to ``/sys/kernel/debug/kcsan``
158 changes the report filtering behaviour. For example, the blacklist feature
159 can be used to silence frequently occurring data races; the whitelist feature
160 can help with reproduction and testing of fixes.
165 Core parameters that affect KCSAN's overall performance and bug detection
166 ability are exposed as kernel command-line arguments whose defaults can also be
167 changed via the corresponding Kconfig options.
169 * ``kcsan.skip_watch`` (``CONFIG_KCSAN_SKIP_WATCH``): Number of per-CPU memory
170 operations to skip, before another watchpoint is set up. Setting up
171 watchpoints more frequently will result in the likelihood of races to be
172 observed to increase. This parameter has the most significant impact on
173 overall system performance and race detection ability.
175 * ``kcsan.udelay_task`` (``CONFIG_KCSAN_UDELAY_TASK``): For tasks, the
176 microsecond delay to stall execution after a watchpoint has been set up.
177 Larger values result in the window in which we may observe a race to
180 * ``kcsan.udelay_interrupt`` (``CONFIG_KCSAN_UDELAY_INTERRUPT``): For
181 interrupts, the microsecond delay to stall execution after a watchpoint has
182 been set up. Interrupts have tighter latency requirements, and their delay
183 should generally be smaller than the one chosen for tasks.
185 They may be tweaked at runtime via ``/sys/module/kcsan/parameters/``.
190 In an execution, two memory accesses form a *data race* if they *conflict*,
191 they happen concurrently in different threads, and at least one of them is a
192 *plain access*; they *conflict* if both access the same memory location, and at
193 least one is a write. For a more thorough discussion and definition, see `"Plain
194 Accesses and Data Races" in the LKMM`_.
196 .. _"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
198 Relationship with the Linux-Kernel Memory Consistency Model (LKMM)
199 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
201 The LKMM defines the propagation and ordering rules of various memory
202 operations, which gives developers the ability to reason about concurrent code.
203 Ultimately this allows to determine the possible executions of concurrent code,
204 and if that code is free from data races.
206 KCSAN is aware of *marked atomic operations* (``READ_ONCE``, ``WRITE_ONCE``,
207 ``atomic_*``, etc.), and a subset of ordering guarantees implied by memory
208 barriers. With ``CONFIG_KCSAN_WEAK_MEMORY=y``, KCSAN models load or store
209 buffering, and can detect missing ``smp_mb()``, ``smp_wmb()``, ``smp_rmb()``,
210 ``smp_store_release()``, and all ``atomic_*`` operations with equivalent
213 Note, KCSAN will not report all data races due to missing memory ordering,
214 specifically where a memory barrier would be required to prohibit subsequent
215 memory operation from reordering before the barrier. Developers should
216 therefore carefully consider the required memory ordering requirements that
219 Race Detection Beyond Data Races
220 --------------------------------
222 For code with complex concurrency design, race-condition bugs may not always
223 manifest as data races. Race conditions occur if concurrently executing
224 operations result in unexpected system behaviour. On the other hand, data races
225 are defined at the C-language level. The following macros can be used to check
226 properties of concurrent code where bugs would not manifest as data races.
228 .. kernel-doc:: include/linux/kcsan-checks.h
229 :functions: ASSERT_EXCLUSIVE_WRITER ASSERT_EXCLUSIVE_WRITER_SCOPED
230 ASSERT_EXCLUSIVE_ACCESS ASSERT_EXCLUSIVE_ACCESS_SCOPED
231 ASSERT_EXCLUSIVE_BITS
233 Implementation Details
234 ----------------------
236 KCSAN relies on observing that two accesses happen concurrently. Crucially, we
237 want to (a) increase the chances of observing races (especially for races that
238 manifest rarely), and (b) be able to actually observe them. We can accomplish
239 (a) by injecting various delays, and (b) by using address watchpoints (or
242 If we deliberately stall a memory access, while we have a watchpoint for its
243 address set up, and then observe the watchpoint to fire, two accesses to the
244 same address just raced. Using hardware watchpoints, this is the approach taken
246 <http://usenix.org/legacy/events/osdi10/tech/full_papers/Erickson.pdf>`_.
247 Unlike DataCollider, KCSAN does not use hardware watchpoints, but instead
248 relies on compiler instrumentation and "soft watchpoints".
250 In KCSAN, watchpoints are implemented using an efficient encoding that stores
251 access type, size, and address in a long; the benefits of using "soft
252 watchpoints" are portability and greater flexibility. KCSAN then relies on the
253 compiler instrumenting plain accesses. For each instrumented plain access:
255 1. Check if a matching watchpoint exists; if yes, and at least one access is a
256 write, then we encountered a racing access.
258 2. Periodically, if no matching watchpoint exists, set up a watchpoint and
259 stall for a small randomized delay.
261 3. Also check the data value before the delay, and re-check the data value
262 after delay; if the values mismatch, we infer a race of unknown origin.
264 To detect data races between plain and marked accesses, KCSAN also annotates
265 marked accesses, but only to check if a watchpoint exists; i.e. KCSAN never
266 sets up a watchpoint on marked accesses. By never setting up watchpoints for
267 marked operations, if all accesses to a variable that is accessed concurrently
268 are properly marked, KCSAN will never trigger a watchpoint and therefore never
274 KCSAN's approach to detecting data races due to missing memory barriers is
275 based on modeling access reordering (with ``CONFIG_KCSAN_WEAK_MEMORY=y``).
276 Each plain memory access for which a watchpoint is set up, is also selected for
277 simulated reordering within the scope of its function (at most 1 in-flight
280 Once an access has been selected for reordering, it is checked along every
281 other access until the end of the function scope. If an appropriate memory
282 barrier is encountered, the access will no longer be considered for simulated
285 When the result of a memory operation should be ordered by a barrier, KCSAN can
286 then detect data races where the conflict only occurs as a result of a missing
287 barrier. Consider the example::
293 WRITE_ONCE(flag, 1); // correct: smp_store_release(&flag, 1)
297 while (!READ_ONCE(flag)); // correct: smp_load_acquire(&flag)
298 ... = x; // data race!
301 When weak memory modeling is enabled, KCSAN can consider ``x`` in ``T1`` for
302 simulated reordering. After the write of ``flag``, ``x`` is again checked for
303 concurrent accesses: because ``T2`` is able to proceed after the write of
304 ``flag``, a data race is detected. With the correct barriers in place, ``x``
305 would not be considered for reordering after the proper release of ``flag``,
306 and no data race would be detected.
308 Deliberate trade-offs in complexity but also practical limitations mean only a
309 subset of data races due to missing memory barriers can be detected. With
310 currently available compiler support, the implementation is limited to modeling
311 the effects of "buffering" (delaying accesses), since the runtime cannot
312 "prefetch" accesses. Also recall that watchpoints are only set up for plain
313 accesses, and the only access type for which KCSAN simulates reordering. This
314 means reordering of marked accesses is not modeled.
316 A consequence of the above is that acquire operations do not require barrier
317 instrumentation (no prefetching). Furthermore, marked accesses introducing
318 address or control dependencies do not require special handling (the marked
319 access cannot be reordered, later dependent accesses cannot be prefetched).
324 1. **Memory Overhead:** The overall memory overhead is only a few MiB
325 depending on configuration. The current implementation uses a small array of
326 longs to encode watchpoint information, which is negligible.
328 2. **Performance Overhead:** KCSAN's runtime aims to be minimal, using an
329 efficient watchpoint encoding that does not require acquiring any shared
330 locks in the fast-path. For kernel boot on a system with 8 CPUs:
332 - 5.0x slow-down with the default KCSAN config;
333 - 2.8x slow-down from runtime fast-path overhead only (set very large
334 ``KCSAN_SKIP_WATCH`` and unset ``KCSAN_SKIP_WATCH_RANDOMIZE``).
336 3. **Annotation Overheads:** Minimal annotations are required outside the KCSAN
337 runtime. As a result, maintenance overheads are minimal as the kernel
340 4. **Detects Racy Writes from Devices:** Due to checking data values upon
341 setting up watchpoints, racy writes from devices can also be detected.
343 5. **Memory Ordering:** KCSAN is aware of only a subset of LKMM ordering rules;
344 this may result in missed data races (false negatives).
346 6. **Analysis Accuracy:** For observed executions, due to using a sampling
347 strategy, the analysis is *unsound* (false negatives possible), but aims to
348 be complete (no false positives).
350 Alternatives Considered
351 -----------------------
353 An alternative data race detection approach for the kernel can be found in the
354 `Kernel Thread Sanitizer (KTSAN) <https://github.com/google/ktsan/wiki>`_.
355 KTSAN is a happens-before data race detector, which explicitly establishes the
356 happens-before order between memory operations, which can then be used to
357 determine data races as defined in `Data Races`_.
359 To build a correct happens-before relation, KTSAN must be aware of all ordering
360 rules of the LKMM and synchronization primitives. Unfortunately, any omission
361 leads to large numbers of false positives, which is especially detrimental in
362 the context of the kernel which includes numerous custom synchronization
363 mechanisms. To track the happens-before relation, KTSAN's implementation
364 requires metadata for each memory location (shadow memory), which for each page
365 corresponds to 4 pages of shadow memory, and can translate into overhead of
366 tens of GiB on a large system.