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 generic_permission / kernfs_refresh_inode
32 write to 0xffff8fee4c40700c of 4 bytes by task 175 on cpu 4:
33 kernfs_refresh_inode+0x70/0x170
34 kernfs_iop_permission+0x4f/0x90
35 inode_permission+0x190/0x200
36 link_path_walk.part.0+0x503/0x8e0
37 path_lookupat.isra.0+0x69/0x4d0
38 filename_lookup+0x136/0x280
39 user_path_at_empty+0x47/0x60
41 __do_sys_newlstat+0x50/0xb0
42 __x64_sys_newlstat+0x37/0x50
43 do_syscall_64+0x85/0x260
44 entry_SYSCALL_64_after_hwframe+0x44/0xa9
46 read to 0xffff8fee4c40700c of 4 bytes by task 166 on cpu 6:
47 generic_permission+0x5b/0x2a0
48 kernfs_iop_permission+0x66/0x90
49 inode_permission+0x190/0x200
50 link_path_walk.part.0+0x503/0x8e0
51 path_lookupat.isra.0+0x69/0x4d0
52 filename_lookup+0x136/0x280
53 user_path_at_empty+0x47/0x60
54 do_faccessat+0x11a/0x390
55 __x64_sys_access+0x3c/0x50
56 do_syscall_64+0x85/0x260
57 entry_SYSCALL_64_after_hwframe+0x44/0xa9
59 Reported by Kernel Concurrency Sanitizer on:
60 CPU: 6 PID: 166 Comm: systemd-journal Not tainted 5.3.0-rc7+ #1
61 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
62 ==================================================================
64 The header of the report provides a short summary of the functions involved in
65 the race. It is followed by the access types and stack traces of the 2 threads
66 involved in the data race.
68 The other less common type of data race report looks like this::
70 ==================================================================
71 BUG: KCSAN: data-race in e1000_clean_rx_irq+0x551/0xb10
73 race at unknown origin, with read to 0xffff933db8a2ae6c of 1 bytes by interrupt on cpu 0:
74 e1000_clean_rx_irq+0x551/0xb10
75 e1000_clean+0x533/0xda0
76 net_rx_action+0x329/0x900
77 __do_softirq+0xdb/0x2db
80 ret_from_intr+0x0/0x18
81 default_idle+0x3f/0x220
82 arch_cpu_idle+0x21/0x30
84 cpu_startup_entry+0x14/0x20
86 arch_call_rest_init+0x13/0x2b
87 start_kernel+0x6db/0x700
89 Reported by Kernel Concurrency Sanitizer on:
90 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.3.0-rc7+ #2
91 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
92 ==================================================================
94 This report is generated where it was not possible to determine the other
95 racing thread, but a race was inferred due to the data value of the watched
96 memory location having changed. These can occur either due to missing
97 instrumentation or e.g. DMA accesses. These reports will only be generated if
98 ``CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=y`` (selected by default).
103 It may be desirable to disable data race detection for specific accesses,
104 functions, compilation units, or entire subsystems. For static blacklisting,
105 the below options are available:
107 * KCSAN understands the ``data_race(expr)`` annotation, which tells KCSAN that
108 any data races due to accesses in ``expr`` should be ignored and resulting
109 behaviour when encountering a data race is deemed safe.
111 * Disabling data race detection for entire functions can be accomplished by
112 using the function attribute ``__no_kcsan``::
118 To dynamically limit for which functions to generate reports, see the
119 `DebugFS interface`_ blacklist/whitelist feature.
121 * To disable data race detection for a particular compilation unit, add to the
124 KCSAN_SANITIZE_file.o := n
126 * To disable data race detection for all compilation units listed in a
127 ``Makefile``, add to the respective ``Makefile``::
131 Furthermore, it is possible to tell KCSAN to show or hide entire classes of
132 data races, depending on preferences. These can be changed via the following
135 * ``CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY``: If enabled and a conflicting write
136 is observed via a watchpoint, but the data value of the memory location was
137 observed to remain unchanged, do not report the data race.
139 * ``CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC``: Assume that plain aligned writes
140 up to word size are atomic by default. Assumes that such writes are not
141 subject to unsafe compiler optimizations resulting in data races. The option
142 causes KCSAN to not report data races due to conflicts where the only plain
143 accesses are aligned writes up to word size.
148 The file ``/sys/kernel/debug/kcsan`` provides the following interface:
150 * Reading ``/sys/kernel/debug/kcsan`` returns various runtime statistics.
152 * Writing ``on`` or ``off`` to ``/sys/kernel/debug/kcsan`` allows turning KCSAN
153 on or off, respectively.
155 * Writing ``!some_func_name`` to ``/sys/kernel/debug/kcsan`` adds
156 ``some_func_name`` to the report filter list, which (by default) blacklists
157 reporting data races where either one of the top stackframes are a function
160 * Writing either ``blacklist`` or ``whitelist`` to ``/sys/kernel/debug/kcsan``
161 changes the report filtering behaviour. For example, the blacklist feature
162 can be used to silence frequently occurring data races; the whitelist feature
163 can help with reproduction and testing of fixes.
168 Core parameters that affect KCSAN's overall performance and bug detection
169 ability are exposed as kernel command-line arguments whose defaults can also be
170 changed via the corresponding Kconfig options.
172 * ``kcsan.skip_watch`` (``CONFIG_KCSAN_SKIP_WATCH``): Number of per-CPU memory
173 operations to skip, before another watchpoint is set up. Setting up
174 watchpoints more frequently will result in the likelihood of races to be
175 observed to increase. This parameter has the most significant impact on
176 overall system performance and race detection ability.
178 * ``kcsan.udelay_task`` (``CONFIG_KCSAN_UDELAY_TASK``): For tasks, the
179 microsecond delay to stall execution after a watchpoint has been set up.
180 Larger values result in the window in which we may observe a race to
183 * ``kcsan.udelay_interrupt`` (``CONFIG_KCSAN_UDELAY_INTERRUPT``): For
184 interrupts, the microsecond delay to stall execution after a watchpoint has
185 been set up. Interrupts have tighter latency requirements, and their delay
186 should generally be smaller than the one chosen for tasks.
188 They may be tweaked at runtime via ``/sys/module/kcsan/parameters/``.
193 In an execution, two memory accesses form a *data race* if they *conflict*,
194 they happen concurrently in different threads, and at least one of them is a
195 *plain access*; they *conflict* if both access the same memory location, and at
196 least one is a write. For a more thorough discussion and definition, see `"Plain
197 Accesses and Data Races" in the LKMM`_.
199 .. _"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
201 Relationship with the Linux-Kernel Memory Consistency Model (LKMM)
202 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
204 The LKMM defines the propagation and ordering rules of various memory
205 operations, which gives developers the ability to reason about concurrent code.
206 Ultimately this allows to determine the possible executions of concurrent code,
207 and if that code is free from data races.
209 KCSAN is aware of *marked atomic operations* (``READ_ONCE``, ``WRITE_ONCE``,
210 ``atomic_*``, etc.), but is oblivious of any ordering guarantees and simply
211 assumes that memory barriers are placed correctly. In other words, KCSAN
212 assumes that as long as a plain access is not observed to race with another
213 conflicting access, memory operations are correctly ordered.
215 This means that KCSAN will not report *potential* data races due to missing
216 memory ordering. Developers should therefore carefully consider the required
217 memory ordering requirements that remain unchecked. If, however, missing
218 memory ordering (that is observable with a particular compiler and
219 architecture) leads to an observable data race (e.g. entering a critical
220 section erroneously), KCSAN would report the resulting data race.
222 Race Detection Beyond Data Races
223 --------------------------------
225 For code with complex concurrency design, race-condition bugs may not always
226 manifest as data races. Race conditions occur if concurrently executing
227 operations result in unexpected system behaviour. On the other hand, data races
228 are defined at the C-language level. The following macros can be used to check
229 properties of concurrent code where bugs would not manifest as data races.
231 .. kernel-doc:: include/linux/kcsan-checks.h
232 :functions: ASSERT_EXCLUSIVE_WRITER ASSERT_EXCLUSIVE_WRITER_SCOPED
233 ASSERT_EXCLUSIVE_ACCESS ASSERT_EXCLUSIVE_ACCESS_SCOPED
234 ASSERT_EXCLUSIVE_BITS
236 Implementation Details
237 ----------------------
239 KCSAN relies on observing that two accesses happen concurrently. Crucially, we
240 want to (a) increase the chances of observing races (especially for races that
241 manifest rarely), and (b) be able to actually observe them. We can accomplish
242 (a) by injecting various delays, and (b) by using address watchpoints (or
245 If we deliberately stall a memory access, while we have a watchpoint for its
246 address set up, and then observe the watchpoint to fire, two accesses to the
247 same address just raced. Using hardware watchpoints, this is the approach taken
249 <http://usenix.org/legacy/events/osdi10/tech/full_papers/Erickson.pdf>`_.
250 Unlike DataCollider, KCSAN does not use hardware watchpoints, but instead
251 relies on compiler instrumentation and "soft watchpoints".
253 In KCSAN, watchpoints are implemented using an efficient encoding that stores
254 access type, size, and address in a long; the benefits of using "soft
255 watchpoints" are portability and greater flexibility. KCSAN then relies on the
256 compiler instrumenting plain accesses. For each instrumented plain access:
258 1. Check if a matching watchpoint exists; if yes, and at least one access is a
259 write, then we encountered a racing access.
261 2. Periodically, if no matching watchpoint exists, set up a watchpoint and
262 stall for a small randomized delay.
264 3. Also check the data value before the delay, and re-check the data value
265 after delay; if the values mismatch, we infer a race of unknown origin.
267 To detect data races between plain and marked accesses, KCSAN also annotates
268 marked accesses, but only to check if a watchpoint exists; i.e. KCSAN never
269 sets up a watchpoint on marked accesses. By never setting up watchpoints for
270 marked operations, if all accesses to a variable that is accessed concurrently
271 are properly marked, KCSAN will never trigger a watchpoint and therefore never
277 1. **Memory Overhead:** The overall memory overhead is only a few MiB
278 depending on configuration. The current implementation uses a small array of
279 longs to encode watchpoint information, which is negligible.
281 2. **Performance Overhead:** KCSAN's runtime aims to be minimal, using an
282 efficient watchpoint encoding that does not require acquiring any shared
283 locks in the fast-path. For kernel boot on a system with 8 CPUs:
285 - 5.0x slow-down with the default KCSAN config;
286 - 2.8x slow-down from runtime fast-path overhead only (set very large
287 ``KCSAN_SKIP_WATCH`` and unset ``KCSAN_SKIP_WATCH_RANDOMIZE``).
289 3. **Annotation Overheads:** Minimal annotations are required outside the KCSAN
290 runtime. As a result, maintenance overheads are minimal as the kernel
293 4. **Detects Racy Writes from Devices:** Due to checking data values upon
294 setting up watchpoints, racy writes from devices can also be detected.
296 5. **Memory Ordering:** KCSAN is *not* explicitly aware of the LKMM's ordering
297 rules; this may result in missed data races (false negatives).
299 6. **Analysis Accuracy:** For observed executions, due to using a sampling
300 strategy, the analysis is *unsound* (false negatives possible), but aims to
301 be complete (no false positives).
303 Alternatives Considered
304 -----------------------
306 An alternative data race detection approach for the kernel can be found in the
307 `Kernel Thread Sanitizer (KTSAN) <https://github.com/google/ktsan/wiki>`_.
308 KTSAN is a happens-before data race detector, which explicitly establishes the
309 happens-before order between memory operations, which can then be used to
310 determine data races as defined in `Data Races`_.
312 To build a correct happens-before relation, KTSAN must be aware of all ordering
313 rules of the LKMM and synchronization primitives. Unfortunately, any omission
314 leads to large numbers of false positives, which is especially detrimental in
315 the context of the kernel which includes numerous custom synchronization
316 mechanisms. To track the happens-before relation, KTSAN's implementation
317 requires metadata for each memory location (shadow memory), which for each page
318 corresponds to 4 pages of shadow memory, and can translate into overhead of
319 tens of GiB on a large system.