Re: Sleeping function called from illegal context...

[William Lee Irwin III <wli@holomorphy.com>]

On Sat, 2002-09-28 at 13:24, John Levon wrote:
>> NMI interrupt handler cannot block so it trylocks against a spinlock
>> instead. Buffer processing code needs to block against concurrent NMI
>> interrupts so takes the spinlock for them. All actual blocks on the
>> spinlock are beneath a down() on another semaphore, so a sleep whilst
>> holding the spinlock won't actually cause deadlock.

On Sat, Sep 28, 2002 at 02:27:44PM -0400, Robert Love wrote:
> If all accesses to the spinlock are taken under a semaphore, then the
> spinlock is not needed (i.e. the down'ed semaphore provides the same
> protection), or am I missing something?
> If this is not the case - e.g. there are other accesses to these locks -
> then you cannot sleep, no?
> I really can think of no case in which it is safe to sleep while holding
> a spinlock or otherwise atomic.  If it is, then the atomicity is not
> needed, sort of by definition.

Actually, though he may be using a spinlock_t, when used this way, it
is not a spinlock, but rather a semaphore-like construct like PG_locked.
Spinlocks include blocking via busywait semantics, which this usage
does not have. It just happens to use the same data type. There are
other interesting abuses of spinlock-like constructs in "advanced"
locks, for instance, in non-sleeping handoff-scheduled queueing locks
(e.g. MCS spinlocks and rwlocks) it's a common idiom for one waiter to
set a "blocked" bit or lock word and then spin on it until another
waiter and/or cpu manipulating the lock clears it.


Bill