RCU Subsystem Details

RCU Subsystem Details

RCU Read-Side Critical Sections

• rcu_read_lock() / rcu_read_unlock() delimit read-side critical sections
• No blocking or sleeping inside classic RCU read sections
• Nesting rcu_read_lock() calls is safe
• Preemption of read sections is allowed under CONFIG_PREEMPT_RCU

Publishing and Reading

• rcu_assign_pointer(p, v): publishes a pointer with release semantics (smp_store_release), ensuring prior initialization is visible to readers
• rcu_dereference(p): loads a pointer with dependency ordering (READ_ONCE), ensuring subsequent accesses through the pointer see published data
• These form a pair: the release in rcu_assign_pointer() orders with the dependency in rcu_dereference()

Grace Period and Reclamation

• synchronize_rcu(): blocks until all pre-existing read-side critical sections complete
• call_rcu(&head, callback): defers callback until after a grace period (non-blocking)
• kfree_rcu(ptr, rhf): shorthand for call_rcu() that calls kfree() in the callback; rhf is the rcu_head field name
• kfree_rcu_mightsleep(ptr): head-less variant that requires no rcu_head field but must be called from sleepable context

RCU Variants

Tasks RCU has no explicit read-side lock — any code that does not voluntarily context-switch is implicitly in a read-side critical section.

Quick Checks

• rcu_read_lock_held() for lockdep debug assertions
• INIT_RCU_HEAD has been removed from the kernel entirely
• rcu_barrier() waits for all pending call_rcu() callbacks to complete (needed at module unload)

RCU-001: Remove Before Reclaim Ordering

Objects must be removed from RCU-protected data structures before calling call_rcu() or synchronize_rcu(). This is because call_rcu() only waits for readers that existed when it was called — it provides no protection against readers that start after the grace period begins. If the object is still linked in the data structure, new readers can find it and access it after it is freed.

The correct sequence:

1. Remove from data structure — prevents new readers from finding the object
2. call_rcu() or synchronize_rcu() — waits for existing readers to finish
3. Free the resource — in the callback or after synchronize_rcu() returns

Use the appropriate RCU-aware removal helpers: hlist_del_rcu(), list_del_rcu(), rhashtable_remove_fast(), etc.

// WRONG — removal after call_rcu causes use-after-free
call_rcu(&obj->rcu, free_callback);

void free_callback(struct rcu_head *rhp) {
    struct obj *obj = container_of(rhp, struct obj, rcu);
    hlist_del_rcu(&obj->node);  // Too late: new readers already found it
    kfree(obj);
}

// CORRECT — remove first, then defer freeing
hlist_del_rcu(&obj->node);         // No new readers can find it
call_rcu(&obj->rcu, free_callback);

void free_callback(struct rcu_head *rhp) {
    struct obj *obj = container_of(rhp, struct obj, rcu);
    kfree(obj);                    // Safe: all prior readers are done
}

REPORT as bugs: Code that calls call_rcu() or kfree_rcu() on an object that is still reachable through an RCU-protected data structure, or code that performs the removal inside the RCU callback rather than before it.

kvfree_call_rcu() / kfree_rcu() Calling Context

kvfree_call_rcu() (called via kfree_rcu/kvfree_rcu macros) is called under raw_spinlock_t (pi_lock in kernel/sched/core.c) and from hardirq context. Adding spinlock_t, local_lock, or local_trylock acquisition in kvfree_call_rcu() or its callees causes a lockdep Invalid wait context warning — !IS_ENABLED(CONFIG_PREEMPT_RT) guards do not prevent this because CONFIG_PROVE_RAW_LOCK_NESTING (default y) checks declared wait-types, not runtime behavior.

Do NOT dismiss this because the same lock types exist elsewhere.