kernel（kernel 效率）

Atomic Context Traps

• spin_lock held = cannot sleep — no kmalloc(GFP_KERNEL), no mutex_lock, no copy_from_user
• Interrupt can take same spinlock — must use spin_lock_irqsave, not plain spin_lock
• rcu_read_lock() section cannot sleep — no blocking calls inside RCU read-side
• might_sleep() annotation — add to functions that may sleep, catches bugs with CONFIG_DEBUG_ATOMIC_SLEEP

Allocation Failures

• GFP_ATOMIC can return NULL — always check, don't assume success
• vmalloc memory not physically contiguous — cannot use for DMA
• kzalloc over kmalloc — uninitialized memory leaks kernel info to userspace
• Allocation in loop risks OOM — preallocate or use memory pool

User Pointer Handling

• copy_from_user returns bytes NOT copied — 0 means success, not failure
• Never use %s with user pointer in printk — kernel crash or info leak
• User memory can change during syscall — copy to kernel buffer, validate the copy
• __user annotation is documentation — doesn't enforce anything, you must use copy functions

Memory Ordering

• READ_ONCE/WRITE_ONCE for lockless shared data — prevents compiler from caching/reordering
• Spinlock release has implicit barrier — but check-then-act patterns still need care
• smp_wmb() before publishing pointer — ensures data visible before pointer is

Module Error Paths

• Init fails midway — must undo everything already done
• Reverse order cleanup — unregister in opposite order of register
• goto err_* pattern standard — cleaner than nested ifs
• Check what's actually initialized — don't free/unregister what wasn't set up

Locking Mistakes

• Same lock acquired twice = deadlock — even in different functions
• Inconsistent lock ordering — document order, acquire in same sequence everywhere
• mutex_trylock returns 1 on success — opposite of pthread_mutex_trylock
• Reader-writer locks rarely worth it — contention overhead usually exceeds benefit