kcsan: Avoid blocking producers in prepare_report()
To avoid deadlock in case watchers can be interrupted, we need to ensure that producers of the struct other_info can never be blocked by an unrelated consumer. (Likely to occur with KCSAN_INTERRUPT_WATCHER.) There are several cases that can lead to this scenario, for example: 1. A watchpoint A was set up by task T1, but interrupted by interrupt I1. Some other thread (task or interrupt) finds watchpoint A consumes it, and sets other_info. Then I1 also finds some unrelated watchpoint B, consumes it, but is blocked because other_info is in use. T1 cannot consume other_info because I1 never returns -> deadlock. 2. A watchpoint A was set up by task T1, but interrupted by interrupt I1, which also sets up a watchpoint B. Some other thread finds watchpoint A, and consumes it and sets up other_info with its information. Similarly some other thread finds watchpoint B and consumes it, but is then blocked because other_info is in use. When I1 continues it sees its watchpoint was consumed, and that it must wait for other_info, which currently contains information to be consumed by T1. However, T1 cannot unblock other_info because I1 never returns -> deadlock. To avoid this, we need to ensure that producers of struct other_info always have a usable other_info entry. This is obviously not the case with only a single instance of struct other_info, as concurrent producers must wait for the entry to be released by some consumer (which may be locked up as illustrated above). While it would be nice if producers could simply call kmalloc() and append their instance of struct other_info to a list, we are very limited in this code path: since KCSAN can instrument the allocators themselves, calling kmalloc() could lead to deadlock or corrupted allocator state. Since producers of the struct other_info will always succeed at try_consume_watchpoint(), preceding the call into kcsan_report(), we know that the particular watchpoint slot cannot simply be reused or consumed by another potential other_info producer. If we move removal of a watchpoint after reporting (by the consumer of struct other_info), we can see a consumed watchpoint as a held lock on elements of other_info, if we create a one-to-one mapping of a watchpoint to an other_info element. Therefore, the simplest solution is to create an array of struct other_info that is as large as the watchpoints array in core.c, and pass the watchpoint index to kcsan_report() for producers and consumers, and change watchpoints to be removed after reporting is done. With a default config on a 64-bit system, the array other_infos consumes ~37KiB. For most systems today this is not a problem. On smaller memory constrained systems, the config value CONFIG_KCSAN_NUM_WATCHPOINTS can be reduced appropriately. Overall, this change is a simplification of the prepare_report() code, and makes some of the checks (such as checking if at least one access is a write) redundant. Tested: $ tools/testing/selftests/rcutorture/bin/kvm.sh \ --cpus 12 --duration 10 --kconfig "CONFIG_DEBUG_INFO=y \ CONFIG_KCSAN=y CONFIG_KCSAN_ASSUME_PLAIN_WRITES_ATOMIC=n \ CONFIG_KCSAN_REPORT_VALUE_CHANGE_ONLY=n \ CONFIG_KCSAN_REPORT_ONCE_IN_MS=100000 CONFIG_KCSAN_VERBOSE=y \ CONFIG_KCSAN_INTERRUPT_WATCHER=y CONFIG_PROVE_LOCKING=y" \ --configs TREE03 => No longer hangs and runs to completion as expected. Reported-by:Paul E. McKenney <paulmck@kernel.org> Signed-off-by:
Marco Elver <elver@google.com> Signed-off-by:
Showing
This diff is collapsed.
Please register or sign in to comment