Re: [DOCUMENTATION] Revised Unreliable Kernel Locking Guide

[Manfred Spraul <manfred@colorfullife.com>]

Hi Rusty,

 From Chapter 4.1:

> |spin_lock_irqsave()| (include/linux/spinlock.h) is a variant which 
> saves whether interrupts were on or off in a flags word, which is 
> passed to |spin_unlock_irqrestore()|. This means that the same code 
> can be used inside an hard irq handler (where interrupts are already 
> off) and in softirqs (where the irq disabling is required).

Interrupts are typically on within the hard irq handler.
spin_lock() is usually ok because an interrupt handler is never 
reentered. Thus if a lock is only accessed from a single irq handler, 
then spin_lock() is the faster approach. That's why many nic drivers use 
spin_lock instead of spin_lock_irqsave() in their irq handlers.
OTHO: if a driver lock is a global resource that is used from different 
irqs, then it must either use _irqsave(), or set SA_INTERRUPT.
Examples: rtc_lock relies on SA_INTERRUPT: it's touched from the rtc irq 
and the timer irq path, and both rtc and timer set SA_INTERRUPT.
I assume ide relies on _irqsave(), but the code is too difficult to follow.

Btw, perhaps you could add the 2nd synchronization approach for 
interrupts: if it's an extremely rare event, then no lock at all in the 
irq handler (no reentrancy guaranteed by the kernel), and 
spin_lock+disable_irq() in the softirq/tasklet. My network drivers use 
that to synchronize packet rx with close.

--
    Manfred