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 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 inexpr
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.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 respectiveMakefile
: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.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 onCONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY=y
. For details please see thekernel/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
oroff
to/sys/kernel/debug/kcsan
allows turning KCSAN on or off, respectively.Writing
!some_func_name
to/sys/kernel/debug/kcsan
addssome_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
orwhitelist
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.
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.
-
ASSERT_EXCLUSIVE_WRITER¶
ASSERT_EXCLUSIVE_WRITER (var)
assert no concurrent writes to var
Parameters
var
variable to assert on
Description
Assert that there are no concurrent writes to var; other readers are allowed. This assertion can be used to specify properties of concurrent code, where violation cannot be detected as a normal data race.
For example, if we only have a single writer, but multiple concurrent readers, to avoid data races, all these accesses must be marked; even concurrent marked writes racing with the single writer are bugs. Unfortunately, due to being marked, they are no longer data races. For cases like these, we can use the macro as follows:
void writer(void) {
spin_lock(&update_foo_lock);
ASSERT_EXCLUSIVE_WRITER(shared_foo);
WRITE_ONCE(shared_foo, ...);
spin_unlock(&update_foo_lock);
}
void reader(void) {
// update_foo_lock does not need to be held!
... = READ_ONCE(shared_foo);
}
Note
ASSERT_EXCLUSIVE_WRITER_SCOPED()
, if applicable, performs more thorough
checking if a clear scope where no concurrent writes are expected exists.
-
ASSERT_EXCLUSIVE_WRITER_SCOPED¶
ASSERT_EXCLUSIVE_WRITER_SCOPED (var)
assert no concurrent writes to var in scope
Parameters
var
variable to assert on
Description
Scoped variant of ASSERT_EXCLUSIVE_WRITER()
.
Assert that there are no concurrent writes to var for the duration of the
scope in which it is introduced. This provides a better way to fully cover
the enclosing scope, compared to multiple ASSERT_EXCLUSIVE_WRITER()
, and
increases the likelihood for KCSAN to detect racing accesses.
For example, it allows finding race-condition bugs that only occur due to state changes within the scope itself:
void writer(void) {
spin_lock(&update_foo_lock);
{
ASSERT_EXCLUSIVE_WRITER_SCOPED(shared_foo);
WRITE_ONCE(shared_foo, 42);
...
// shared_foo should still be 42 here!
}
spin_unlock(&update_foo_lock);
}
void buggy(void) {
if (READ_ONCE(shared_foo) == 42)
WRITE_ONCE(shared_foo, 1); // bug!
}
-
ASSERT_EXCLUSIVE_ACCESS¶
ASSERT_EXCLUSIVE_ACCESS (var)
assert no concurrent accesses to var
Parameters
var
variable to assert on
Description
Assert that there are no concurrent accesses to var (no readers nor writers). This assertion can be used to specify properties of concurrent code, where violation cannot be detected as a normal data race.
For example, where exclusive access is expected after determining no other users of an object are left, but the object is not actually freed. We can check that this property actually holds as follows:
if (refcount_dec_and_test(&obj->refcnt)) {
ASSERT_EXCLUSIVE_ACCESS(*obj);
do_some_cleanup(obj);
release_for_reuse(obj);
}
ASSERT_EXCLUSIVE_ACCESS_SCOPED()
, if applicable, performs more thorough checking if a clear scope where no concurrent accesses are expected exists.For cases where the object is freed, KASAN is a better fit to detect use-after-free bugs.
Note
-
ASSERT_EXCLUSIVE_ACCESS_SCOPED¶
ASSERT_EXCLUSIVE_ACCESS_SCOPED (var)
assert no concurrent accesses to var in scope
Parameters
var
variable to assert on
Description
Scoped variant of ASSERT_EXCLUSIVE_ACCESS()
.
Assert that there are no concurrent accesses to var (no readers nor writers)
for the entire duration of the scope in which it is introduced. This provides
a better way to fully cover the enclosing scope, compared to multiple
ASSERT_EXCLUSIVE_ACCESS()
, and increases the likelihood for KCSAN to detect
racing accesses.
-
ASSERT_EXCLUSIVE_BITS¶
ASSERT_EXCLUSIVE_BITS (var, mask)
assert no concurrent writes to subset of bits in var
Parameters
var
variable to assert on
mask
only check for modifications to bits set in mask
Description
Bit-granular variant of ASSERT_EXCLUSIVE_WRITER()
.
Assert that there are no concurrent writes to a subset of bits in var; concurrent readers are permitted. This assertion captures more detailed bit-level properties, compared to the other (word granularity) assertions. Only the bits set in mask are checked for concurrent modifications, while ignoring the remaining bits, i.e. concurrent writes (or reads) to ~mask bits are ignored.
Use this for variables, where some bits must not be modified concurrently, yet other bits are expected to be modified concurrently.
For example, variables where, after initialization, some bits are read-only, but other bits may still be modified concurrently. A reader may wish to assert that this is true as follows:
ASSERT_EXCLUSIVE_BITS(flags, READ_ONLY_MASK);
foo = (READ_ONCE(flags) & READ_ONLY_MASK) >> READ_ONLY_SHIFT;
ASSERT_EXCLUSIVE_BITS(flags, READ_ONLY_MASK);
foo = (flags & READ_ONLY_MASK) >> READ_ONLY_SHIFT;
Another example, where this may be used, is when certain bits of var may only be modified when holding the appropriate lock, but other bits may still be modified concurrently. Writers, where other bits may change concurrently, could use the assertion as follows:
spin_lock(&foo_lock);
ASSERT_EXCLUSIVE_BITS(flags, FOO_MASK);
old_flags = flags;
new_flags = (old_flags & ~FOO_MASK) | (new_foo << FOO_SHIFT);
if (cmpxchg(&flags, old_flags, new_flags) != old_flags) { ... }
spin_unlock(&foo_lock);
Note
The access that immediately follows ASSERT_EXCLUSIVE_BITS()
is assumed
to access the masked bits only, and KCSAN optimistically assumes it is
therefore safe, even in the presence of data races, and marking it with
READ_ONCE() is optional from KCSAN’s point-of-view. We caution, however, that
it may still be advisable to do so, since we cannot reason about all compiler
optimizations when it comes to bit manipulations (on the reader and writer
side). If you are sure nothing can go wrong, we can write the above simply
as:
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. 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:
Check if a matching watchpoint exists; if yes, and at least one access is a write, then we encountered a racing access.
Periodically, if no matching watchpoint exists, set up a watchpoint and stall for a small randomized delay.
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¶
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.
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 unsetKCSAN_SKIP_WATCH_RANDOMIZE
).
Annotation Overheads: Minimal annotations are required outside the KCSAN runtime. As a result, maintenance overheads are minimal as the kernel evolves.
Detects Racy Writes from Devices: Due to checking data values upon setting up watchpoints, racy writes from devices can also be detected.
Memory Ordering: KCSAN is aware of only a subset of LKMM ordering rules; this may result in missed data races (false negatives).
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). 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.