1 The Kernel Concurrency Sanitizer (KCSAN)
2 ========================================
4 The Kernel Concurrency Sanitizer (KCSAN) is a dynamic race detector, which
5 relies on compile-time instrumentation, and uses a watchpoint-based sampling
6 approach to detect races. KCSAN's primary purpose is to detect `data races`_.
11 KCSAN is supported by both GCC and Clang. With GCC we require version 11 or
12 later, and with Clang also require version 11 or later.
14 To enable KCSAN configure the kernel with::
18 KCSAN provides several other configuration options to customize behaviour (see
19 the respective help text in ``lib/Kconfig.kcsan`` for more info).
24 A typical data race report looks like this::
26 ==================================================================
27 BUG: KCSAN: data-race in generic_permission / kernfs_refresh_inode
29 write to 0xffff8fee4c40700c of 4 bytes by task 175 on cpu 4:
30 kernfs_refresh_inode+0x70/0x170
31 kernfs_iop_permission+0x4f/0x90
32 inode_permission+0x190/0x200
33 link_path_walk.part.0+0x503/0x8e0
34 path_lookupat.isra.0+0x69/0x4d0
35 filename_lookup+0x136/0x280
36 user_path_at_empty+0x47/0x60
38 __do_sys_newlstat+0x50/0xb0
39 __x64_sys_newlstat+0x37/0x50
40 do_syscall_64+0x85/0x260
41 entry_SYSCALL_64_after_hwframe+0x44/0xa9
43 read to 0xffff8fee4c40700c of 4 bytes by task 166 on cpu 6:
44 generic_permission+0x5b/0x2a0
45 kernfs_iop_permission+0x66/0x90
46 inode_permission+0x190/0x200
47 link_path_walk.part.0+0x503/0x8e0
48 path_lookupat.isra.0+0x69/0x4d0
49 filename_lookup+0x136/0x280
50 user_path_at_empty+0x47/0x60
51 do_faccessat+0x11a/0x390
52 __x64_sys_access+0x3c/0x50
53 do_syscall_64+0x85/0x260
54 entry_SYSCALL_64_after_hwframe+0x44/0xa9
56 Reported by Kernel Concurrency Sanitizer on:
57 CPU: 6 PID: 166 Comm: systemd-journal Not tainted 5.3.0-rc7+ #1
58 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
59 ==================================================================
61 The header of the report provides a short summary of the functions involved in
62 the race. It is followed by the access types and stack traces of the 2 threads
63 involved in the data race.
65 The other less common type of data race report looks like this::
67 ==================================================================
68 BUG: KCSAN: data-race in e1000_clean_rx_irq+0x551/0xb10
70 race at unknown origin, with read to 0xffff933db8a2ae6c of 1 bytes by interrupt on cpu 0:
71 e1000_clean_rx_irq+0x551/0xb10
72 e1000_clean+0x533/0xda0
73 net_rx_action+0x329/0x900
74 __do_softirq+0xdb/0x2db
77 ret_from_intr+0x0/0x18
78 default_idle+0x3f/0x220
79 arch_cpu_idle+0x21/0x30
81 cpu_startup_entry+0x14/0x20
83 arch_call_rest_init+0x13/0x2b
84 start_kernel+0x6db/0x700
86 Reported by Kernel Concurrency Sanitizer on:
87 CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.3.0-rc7+ #2
88 Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.12.0-1 04/01/2014
89 ==================================================================
91 This report is generated where it was not possible to determine the other
92 racing thread, but a race was inferred due to the data value of the watched
93 memory location having changed. These can occur either due to missing
94 instrumentation or e.g. DMA accesses. These reports will only be generated if
95 ``CONFIG_KCSAN_REPORT_RACE_UNKNOWN_ORIGIN=y`` (selected by default).
100 It may be desirable to disable data race detection for specific accesses,
101 functions, compilation units, or entire subsystems. For static blacklisting,
102 the below options are available:
104 * KCSAN understands the ``data_race(expr)`` annotation, which tells KCSAN that
105 any data races due to accesses in ``expr`` should be ignored and resulting
106 behaviour when encountering a data race is deemed safe.
108 * Disabling data race detection for entire functions can be accomplished by
109 using the function attribute ``__no_kcsan``::
115 To dynamically limit for which functions to generate reports, see the
116 `DebugFS interface`_ blacklist/whitelist feature.
118 * To disable data race detection for a particular compilation unit, add to the
121 KCSAN_SANITIZE_file.o := n
123 * To disable data race detection for all compilation units listed in a
124 ``Makefile``, add to the respective ``Makefile``::
128 Furthermore, it is possible to tell KCSAN to show or hide entire classes of
129 data races, depending on preferences. These can be changed via the following
132 * ``CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY``: If enabled and a conflicting write
133 is observed via a watchpoint, but the data value of the memory location was
134 observed to remain unchanged, do not report the data race.
136 * ``CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC``: Assume that plain aligned writes
137 up to word size are atomic by default. Assumes that such writes are not
138 subject to unsafe compiler optimizations resulting in data races. The option
139 causes KCSAN to not report data races due to conflicts where the only plain
140 accesses are aligned writes up to word size.
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.), but is oblivious of any ordering guarantees and simply
208 assumes that memory barriers are placed correctly. In other words, KCSAN
209 assumes that as long as a plain access is not observed to race with another
210 conflicting access, memory operations are correctly ordered.
212 This means that KCSAN will not report *potential* data races due to missing
213 memory ordering. Developers should therefore carefully consider the required
214 memory ordering requirements that remain unchecked. If, however, missing
215 memory ordering (that is observable with a particular compiler and
216 architecture) leads to an observable data race (e.g. entering a critical
217 section erroneously), KCSAN would report the resulting data race.
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 1. **Memory Overhead:** The overall memory overhead is only a few MiB
275 depending on configuration. The current implementation uses a small array of
276 longs to encode watchpoint information, which is negligible.
278 2. **Performance Overhead:** KCSAN's runtime aims to be minimal, using an
279 efficient watchpoint encoding that does not require acquiring any shared
280 locks in the fast-path. For kernel boot on a system with 8 CPUs:
282 - 5.0x slow-down with the default KCSAN config;
283 - 2.8x slow-down from runtime fast-path overhead only (set very large
284 ``KCSAN_SKIP_WATCH`` and unset ``KCSAN_SKIP_WATCH_RANDOMIZE``).
286 3. **Annotation Overheads:** Minimal annotations are required outside the KCSAN
287 runtime. As a result, maintenance overheads are minimal as the kernel
290 4. **Detects Racy Writes from Devices:** Due to checking data values upon
291 setting up watchpoints, racy writes from devices can also be detected.
293 5. **Memory Ordering:** KCSAN is *not* explicitly aware of the LKMM's ordering
294 rules; this may result in missed data races (false negatives).
296 6. **Analysis Accuracy:** For observed executions, due to using a sampling
297 strategy, the analysis is *unsound* (false negatives possible), but aims to
298 be complete (no false positives).
300 Alternatives Considered
301 -----------------------
303 An alternative data race detection approach for the kernel can be found in the
304 `Kernel Thread Sanitizer (KTSAN) <https://github.com/google/ktsan/wiki>`_.
305 KTSAN is a happens-before data race detector, which explicitly establishes the
306 happens-before order between memory operations, which can then be used to
307 determine data races as defined in `Data Races`_.
309 To build a correct happens-before relation, KTSAN must be aware of all ordering
310 rules of the LKMM and synchronization primitives. Unfortunately, any omission
311 leads to large numbers of false positives, which is especially detrimental in
312 the context of the kernel which includes numerous custom synchronization
313 mechanisms. To track the happens-before relation, KTSAN's implementation
314 requires metadata for each memory location (shadow memory), which for each page
315 corresponds to 4 pages of shadow memory, and can translate into overhead of
316 tens of GiB on a large system.