| Message ID | 20250206181711.1902989-1-elver@google.com |
|---|---|
| Headers | show |
| Series | Compiler-Based Capability- and Locking-Analysis | expand |
On Thu, Feb 06, 2025 at 07:10:05PM +0100, Marco Elver wrote: > extern int __must_check mutex_lock_interruptible_nested(struct mutex *lock, > + unsigned int subclass) __cond_acquires(0, lock); > extern int __must_check mutex_lock_killable_nested(struct mutex *lock, > + unsigned int subclass) __cond_acquires(0, lock); > +extern int __must_check mutex_lock_interruptible(struct mutex *lock) __cond_acquires(0, lock); > +extern int __must_check mutex_lock_killable(struct mutex *lock) __cond_acquires(0, lock); > +extern int mutex_trylock(struct mutex *lock) __cond_acquires(1, lock); > +extern int atomic_dec_and_mutex_lock(atomic_t *cnt, struct mutex *lock) __cond_acquires(1, lock); So this form is *MUCH* saner than what we currently have. Can we please fix up all the existing __cond_lock() code too?
On Thu, 20 Feb 2025 at 23:00, Paul E. McKenney <paulmck@kernel.org> wrote: > > On Thu, Feb 06, 2025 at 07:10:09PM +0100, Marco Elver wrote: > > Improve the existing annotations to properly support Clang's capability > > analysis. > > > > The old annotations distinguished between RCU, RCU_BH, and RCU_SCHED. > > However, it does not make sense to acquire rcu_read_lock_bh() after > > rcu_read_lock() - annotate the _bh() and _sched() variants to also > > acquire 'RCU', so that Clang (and also Sparse) can warn about it. > > You lost me on this one. What breaks if rcu_read_lock_bh() is invoked > while rcu_read_lock() is in effect? I thought something like this does not make sense: rcu_read_lock_bh(); .. rcu_read_lock(); .. rcu_read_unlock(); .. rcu_read_unlock_bh(); However, the inverse may well be something we might find somewhere in the kernel? Another problem was that if we want to indicate that "RCU" read lock is held, then we should just be able to write "__must_hold_shared(RCU)", and it shouldn't matter if rcu_read_lock() or rcu_read_lock_bh() was used. Previously each of them acquired their own capability "RCU" and "RCU_BH" respectively. But rather, we're dealing with one acquiring a superset of the other, and expressing that is also what I attempted to solve. Let me rethink this... Thanks, -- Marco
On Thu, Feb 20, 2025 at 11:11:04PM +0100, Marco Elver wrote: > On Thu, 20 Feb 2025 at 23:00, Paul E. McKenney <paulmck@kernel.org> wrote: > > > > On Thu, Feb 06, 2025 at 07:10:09PM +0100, Marco Elver wrote: > > > Improve the existing annotations to properly support Clang's capability > > > analysis. > > > > > > The old annotations distinguished between RCU, RCU_BH, and RCU_SCHED. > > > However, it does not make sense to acquire rcu_read_lock_bh() after > > > rcu_read_lock() - annotate the _bh() and _sched() variants to also > > > acquire 'RCU', so that Clang (and also Sparse) can warn about it. > > > > You lost me on this one. What breaks if rcu_read_lock_bh() is invoked > > while rcu_read_lock() is in effect? > > I thought something like this does not make sense: > > rcu_read_lock_bh(); > .. > rcu_read_lock(); > .. > rcu_read_unlock(); > .. > rcu_read_unlock_bh(); If you have the choice, it is often better to do the rcu_read_lock() first and the rcu_read_lock_bh() second. > However, the inverse may well be something we might find somewhere in > the kernel? Suppose that one function walks an RCU-protected list, calling some function from some other subsystem on each element. Suppose that each element has another RCU protected list. It would be good if the two subsystems could just choose their desired flavor of RCU reader, without having to know about each other. > Another problem was that if we want to indicate that "RCU" read lock > is held, then we should just be able to write > "__must_hold_shared(RCU)", and it shouldn't matter if rcu_read_lock() > or rcu_read_lock_bh() was used. Previously each of them acquired their > own capability "RCU" and "RCU_BH" respectively. But rather, we're > dealing with one acquiring a superset of the other, and expressing > that is also what I attempted to solve. > Let me rethink this... Would it work to have just one sort of RCU reader, relying on a separate BH-disable capability for the additional semantics of rcu_read_lock_bh()? Thanx, Paul
[ Note: Bart and I had concurrently been working on bringing Clang's -Wthread-safety to the kernel: https://lore.kernel.org/all/20250206175114.1974171-1-bvanassche@acm.org/ Having both RFCs out should hopefully provide a good picture on these design points and trade-offs - the approaches differ significantly. ] Capability analysis is a C language extension, which enables statically checking that user-definable "capabilities" are acquired and released where required. An obvious application is lock-safety checking for the kernel's various synchronization primitives (each of which represents a "capability"), and checking that locking rules are not violated. Clang originally called the feature "Thread Safety Analysis" [1], with some terminology still using the thread-safety-analysis-only names. This was later changed and the feature became more flexible, gaining the ability to define custom "capabilities". Its foundations can be found in "capability systems", used to specify the permissibility of operations to depend on some capability being held (or not held). [1] https://clang.llvm.org/docs/ThreadSafetyAnalysis.html [2] https://www.cs.cornell.edu/talc/papers/capabilities.pdf Because the feature is not just able to express capabilities related to synchronization primitives, the naming chosen for the kernel departs from Clang's initial "Thread Safety" nomenclature and refers to the feature as "Capability Analysis" to avoid confusion. The implementation still makes references to the older terminology in some places, such as `-Wthread-safety` being the warning enabled option that also still appears in diagnostic messages. Enabling capability analysis can be seen as enabling a dialect of Linux C with a Capability System. Additional details can be found in the added kernel-doc documentation. === Development Approach === Prior art exists in the form of Sparse's context tracking. Locking annotations on functions exist, so the concept of analyzing locking rules is not foreign to the kernel's codebase. However, Clang's analysis is more complete vs. Sparse's, with the typical trade-offs in static analysis: improved completeness is sacrificed for more possible false positives or additional annotations required by the programmer. Numerous options exist to disable or opt out certain code from analysis. This series aims to retain compatibility with Sparse, which can provide tree-wide analysis of a subset of the capability analysis introduced. For the most part, the new (and old) keywords used for annotations are shared between Sparse and Clang. One big question is how to enable this feature, given we end up with a new dialect of C - 2 approaches have been considered: A. Tree-wide all-or-nothing approach. This approach requires tree-wide changes, adding annotations or selective opt-outs. Making additional primitives capability-enabled increases churn, esp. where maintainers are unaware of the feature's existence and how to use it. Because we can't change the programming language (even if from one C dialect to another) of the kernel overnight, a different approach might cause less friction. B. A selective, incremental, and much less intrusive approach. Maintainers of subsystems opt in their modules or directories into "capability analysis" (via Makefile): CAPABILITY_ANALYSIS_foo.o := y # foo.o only CAPABILITY_ANALYSIS := y # all TUs Most (eventually all) synchronization primitives and more capabilities (including ones that could track "irq disabled", "preemption" disabled, etc.) could be supported. The approach taken by this series if B. This ensures that only subsystems where maintainers are willing to deal with any warnings one way or another. Introducing the feature can be done incrementally, without large tree-wide changes and adding numerous opt-outs and annotations to the majority of code. === Initial Uses === With this initial series, the following synchronization primitives are supported: `raw_spinlock_t`, `spinlock_t`, `rwlock_t`, `mutex`, `seqlock_t`, `bit_spinlock`, RCU, SRCU (`srcu_struct`), `rw_semaphore`, `local_lock_t`. As an initial proof-of-concept, this series also enables capability analysis for the following subsystems: kfence, kcov, stackdepot, rhashtable. (Those subsystems were chosen because I am familiar with their locking rules; rhashtable was chosen semi-randomly as a test because it combines a bunch of things: RCU, mutex, bit_spinlock.) The initial benefits are static detection of violations of locking rules. As more capabilities are added, we would see more static checking beyond what regular C can provide, all while remaining easy (read quick) to use via the Clang compiler. === Appendix === The following pending Clang patch is recommended, but not a strong dependency: https://github.com/llvm/llvm-project/pull/123063 This series is also available at this Git tree: https://git.kernel.org/pub/scm/linux/kernel/git/melver/linux.git/log/?h=cap-analysis Marco Elver (24): compiler_types: Move lock checking attributes to compiler-capability-analysis.h compiler-capability-analysis: Rename __cond_lock() to __cond_acquire() compiler-capability-analysis: Add infrastructure for Clang's capability analysis compiler-capability-analysis: Add test stub Documentation: Add documentation for Compiler-Based Capability Analysis checkpatch: Warn about capability_unsafe() without comment cleanup: Basic compatibility with capability analysis lockdep: Annotate lockdep assertions for capability analysis locking/rwlock, spinlock: Support Clang's capability analysis compiler-capability-analysis: Change __cond_acquires to take return value locking/mutex: Support Clang's capability analysis locking/seqlock: Support Clang's capability analysis bit_spinlock: Include missing <asm/processor.h> bit_spinlock: Support Clang's capability analysis rcu: Support Clang's capability analysis srcu: Support Clang's capability analysis kref: Add capability-analysis annotations locking/rwsem: Support Clang's capability analysis locking/local_lock: Support Clang's capability analysis debugfs: Make debugfs_cancellation a capability struct kfence: Enable capability analysis kcov: Enable capability analysis stackdepot: Enable capability analysis rhashtable: Enable capability analysis .../dev-tools/capability-analysis.rst | 149 ++++++ Documentation/dev-tools/index.rst | 1 + Documentation/dev-tools/sparse.rst | 4 + Makefile | 1 + .../net/wireless/intel/iwlwifi/iwl-trans.h | 2 +- .../wireless/intel/iwlwifi/pcie/internal.h | 2 +- fs/dlm/lock.c | 2 +- include/linux/bit_spinlock.h | 24 +- include/linux/cleanup.h | 18 +- include/linux/compiler-capability-analysis.h | 407 +++++++++++++++ include/linux/compiler_types.h | 18 +- include/linux/debugfs.h | 12 +- include/linux/kref.h | 2 + include/linux/list_bl.h | 2 + include/linux/local_lock.h | 18 +- include/linux/local_lock_internal.h | 41 +- include/linux/lockdep.h | 12 +- include/linux/mm.h | 6 +- include/linux/mutex.h | 29 +- include/linux/mutex_types.h | 4 +- include/linux/rcupdate.h | 73 ++- include/linux/refcount.h | 6 +- include/linux/rhashtable.h | 14 +- include/linux/rwlock.h | 27 +- include/linux/rwlock_api_smp.h | 29 +- include/linux/rwlock_rt.h | 37 +- include/linux/rwlock_types.h | 10 +- include/linux/rwsem.h | 56 +- include/linux/sched/signal.h | 2 +- include/linux/seqlock.h | 24 + include/linux/seqlock_types.h | 5 +- include/linux/spinlock.h | 61 ++- include/linux/spinlock_api_smp.h | 14 +- include/linux/spinlock_api_up.h | 71 +-- include/linux/spinlock_rt.h | 27 +- include/linux/spinlock_types.h | 10 +- include/linux/spinlock_types_raw.h | 5 +- include/linux/srcu.h | 61 ++- kernel/Makefile | 2 + kernel/kcov.c | 40 +- kernel/time/posix-timers.c | 2 +- lib/Kconfig.debug | 43 ++ lib/Makefile | 6 + lib/rhashtable.c | 12 +- lib/stackdepot.c | 24 +- lib/test_capability-analysis.c | 481 ++++++++++++++++++ mm/kfence/Makefile | 2 + mm/kfence/core.c | 24 +- mm/kfence/kfence.h | 18 +- mm/kfence/kfence_test.c | 4 + mm/kfence/report.c | 8 +- net/ipv4/tcp_sigpool.c | 2 +- scripts/Makefile.capability-analysis | 5 + scripts/Makefile.lib | 10 + scripts/checkpatch.pl | 8 + tools/include/linux/compiler_types.h | 4 +- 56 files changed, 1682 insertions(+), 299 deletions(-) create mode 100644 Documentation/dev-tools/capability-analysis.rst create mode 100644 include/linux/compiler-capability-analysis.h create mode 100644 lib/test_capability-analysis.c create mode 100644 scripts/Makefile.capability-analysis