forked from Minki/linux
d8fd74d35a
Add rules to ignore data-racy reads with only 1-bit value changes. Details about the rules are captured in comments in kernel/kcsan/permissive.h. More background follows. While investigating a number of data races, we've encountered data-racy accesses on flags variables to be very common. The typical pattern is a reader masking all but one bit, and/or the writer setting/clearing only 1 bit (current->flags being a frequently encountered case; more examples in mm/sl[au]b.c, which disable KCSAN for this reason). Since these types of data-racy accesses are common (with the assumption they are intentional and hard to miscompile) having the option (with CONFIG_KCSAN_PERMISSIVE=y) to filter them will avoid forcing everyone to mark them, and deliberately left to preference at this time. One important motivation for having this option built-in is to move closer to being able to enable KCSAN on CI systems or for testers wishing to test the whole kernel, while more easily filtering less interesting data races with higher probability. For the implementation, we considered several alternatives, but had one major requirement: that the rules be kept together with the Linux-kernel tree. Adding them to the compiler would preclude us from making changes quickly; if the rules require tweaks, having them part of the compiler requires waiting another ~1 year for the next release -- that's not realistic. We are left with the following options: 1. Maintain compiler plugins as part of the kernel-tree that removes instrumentation for some accesses (e.g. plain-& with 1-bit mask). The analysis would be reader-side focused, as no assumption can be made about racing writers. Because it seems unrealistic to maintain 2 plugins, one for LLVM and GCC, we would likely pick LLVM. Furthermore, no kernel infrastructure exists to maintain LLVM plugins, and the build-system implications and maintenance overheads do not look great (historically, plugins written against old LLVM APIs are not guaranteed to work with newer LLVM APIs). 2. Find a set of rules that can be expressed in terms of observed value changes, and make it part of the KCSAN runtime. The analysis is writer-side focused, given we rely on observed value changes. The approach taken here is (2). While a complete approach requires both (1) and (2), experiments show that the majority of data races involving trivial bit operations on flags variables can be removed with (2) alone. It goes without saying that the filtering of data races using (1) or (2) does _not_ guarantee they are safe! Therefore, limiting ourselves to (2) for now is the conservative choice for setups that wish to enable CONFIG_KCSAN_PERMISSIVE=y. Signed-off-by: Marco Elver <elver@google.com> Acked-by: Mark Rutland <mark.rutland@arm.com> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
95 lines
2.7 KiB
C
95 lines
2.7 KiB
C
/* SPDX-License-Identifier: GPL-2.0 */
|
|
/*
|
|
* Special rules for ignoring entire classes of data-racy memory accesses. None
|
|
* of the rules here imply that such data races are generally safe!
|
|
*
|
|
* All rules in this file can be configured via CONFIG_KCSAN_PERMISSIVE. Keep
|
|
* them separate from core code to make it easier to audit.
|
|
*
|
|
* Copyright (C) 2019, Google LLC.
|
|
*/
|
|
|
|
#ifndef _KERNEL_KCSAN_PERMISSIVE_H
|
|
#define _KERNEL_KCSAN_PERMISSIVE_H
|
|
|
|
#include <linux/bitops.h>
|
|
#include <linux/sched.h>
|
|
#include <linux/types.h>
|
|
|
|
/*
|
|
* Access ignore rules based on address.
|
|
*/
|
|
static __always_inline bool kcsan_ignore_address(const volatile void *ptr)
|
|
{
|
|
if (!IS_ENABLED(CONFIG_KCSAN_PERMISSIVE))
|
|
return false;
|
|
|
|
/*
|
|
* Data-racy bitops on current->flags are too common, ignore completely
|
|
* for now.
|
|
*/
|
|
return ptr == ¤t->flags;
|
|
}
|
|
|
|
/*
|
|
* Data race ignore rules based on access type and value change patterns.
|
|
*/
|
|
static bool
|
|
kcsan_ignore_data_race(size_t size, int type, u64 old, u64 new, u64 diff)
|
|
{
|
|
if (!IS_ENABLED(CONFIG_KCSAN_PERMISSIVE))
|
|
return false;
|
|
|
|
/*
|
|
* Rules here are only for plain read accesses, so that we still report
|
|
* data races between plain read-write accesses.
|
|
*/
|
|
if (type || size > sizeof(long))
|
|
return false;
|
|
|
|
/*
|
|
* A common pattern is checking/setting just 1 bit in a variable; for
|
|
* example:
|
|
*
|
|
* if (flags & SOME_FLAG) { ... }
|
|
*
|
|
* and elsewhere flags is updated concurrently:
|
|
*
|
|
* flags |= SOME_OTHER_FLAG; // just 1 bit
|
|
*
|
|
* While it is still recommended that such accesses be marked
|
|
* appropriately, in many cases these types of data races are so common
|
|
* that marking them all is often unrealistic and left to maintainer
|
|
* preference.
|
|
*
|
|
* The assumption in all cases is that with all known compiler
|
|
* optimizations (including those that tear accesses), because no more
|
|
* than 1 bit changed, the plain accesses are safe despite the presence
|
|
* of data races.
|
|
*
|
|
* The rules here will ignore the data races if we observe no more than
|
|
* 1 bit changed.
|
|
*
|
|
* Of course many operations can effecively change just 1 bit, but the
|
|
* general assuption that data races involving 1-bit changes can be
|
|
* tolerated still applies.
|
|
*
|
|
* And in case a true bug is missed, the bug likely manifests as a
|
|
* reportable data race elsewhere.
|
|
*/
|
|
if (hweight64(diff) == 1) {
|
|
/*
|
|
* Exception: Report data races where the values look like
|
|
* ordinary booleans (one of them was 0 and the 0th bit was
|
|
* changed) More often than not, they come with interesting
|
|
* memory ordering requirements, so let's report them.
|
|
*/
|
|
if (!((!old || !new) && diff == 1))
|
|
return true;
|
|
}
|
|
|
|
return false;
|
|
}
|
|
|
|
#endif /* _KERNEL_KCSAN_PERMISSIVE_H */
|