Re: [PATCH v6 6/6] lib/dlock-list: Provide IRQ-safe APIs

From: Waiman Long
Date: Thu Oct 05 2017 - 09:41:03 EST

On 10/04/2017 06:46 PM, Davidlohr Bueso wrote:
> On Wed, 04 Oct 2017, Waiman Long wrote:
>> To enable the use of dlock-list in an interrupt handler, the following
>> new APIs are provided for a irqsafe dlock-list:
>> - void dlock_list_unlock_irqsafe(struct dlock_list_iter *)
>> - void dlock_list_relock_irqsafe(struct dlock_list_iter *)
>> - void dlock_list_add_irqsafe(struct dlock_list_node *,
>> struct dlock_list_head *);
>> - void dlock_lists_add_irqsafe(struct dlock_list_node *,
>> struct dlock_list_heads *)
>> - void dlock_lists_del_irqsafe(struct dlock_list_node *)
>> New macros for irqsafe dlock-list:
>> - dlist_for_each_entry_irqsafe(pos, iter, member)
>> - dlist_for_each_entry_safe_irqsafe(pos, n, iter, member)
> Instead of adding more calls to the api, could we not just use the
> irqsave/restore as part of the regular api?
> Thanks,
> Davidlohr

The irqsave/restore spinlock calls are more expensive in term of
performance. I think the spin_lock_irqrestore() is especially bad in a
VM as it probably causes a VMexit. So I try to avoid them unless it is
absolutely necessary.

Another alternative is to specify the dlock-list type at allocation time
and use either regular spinlock calls or irqsave/restore calls
accordingly. That will add a bit of overhead for users that don't need
irq safety, but much less than using irqsave/restore for all. I was
using that approach originally, but opt for the current solution for
performance reason. I can revert back to my original approach and send
out an updated patch.