Re: [KVM-RFC PATCH 1/2] eventfd: add an explicit srcu based notifierinterface

From: Gregory Haskins
Date: Tue Jun 16 2009 - 10:22:41 EST


[Adding Ingo]

Gregory Haskins wrote:
> Michael S. Tsirkin wrote:
>
>> On Mon, Jun 15, 2009 at 10:29:56PM -0400, Gregory Haskins wrote:
>>
>>
>>> irqfd and its underlying implementation, eventfd, currently utilize
>>> the embedded wait-queue in eventfd for signal notification. The nice thing
>>> about this design decision is that it re-uses the existing
>>> eventfd/wait-queue code and it generally works well....with several
>>> limitations.
>>>
>>> One of the limitations is that notification callbacks are always called
>>> inside a spin_lock_irqsave critical section. Another limitation is
>>> that it is very difficult to build a system that can recieve release
>>> notification without being racy.
>>>
>>> Therefore, we introduce a new registration interface that is SRCU based
>>> instead of wait-queue based, and implement the internal wait-queue
>>> infrastructure in terms of this new interface. We then convert irqfd
>>> to use this new interface instead of the existing wait-queue code.
>>>
>>> The end result is that we now have the opportunity to run the interrupt
>>> injection code serially to the callback (when the signal is raised from
>>> process-context, at least) instead of always deferring the injection to a
>>> work-queue.
>>>
>>> Signed-off-by: Gregory Haskins <ghaskins@xxxxxxxxxx>
>>> CC: Paul E. McKenney <paulmck@xxxxxxxxxxxxxxxxxx>
>>> CC: Davide Libenzi <davidel@xxxxxxxxxxxxxxx>
>>> ---
>>>
>>> fs/eventfd.c | 115 +++++++++++++++++++++++++++++++++++++++++++----
>>> include/linux/eventfd.h | 30 ++++++++++++
>>> virt/kvm/eventfd.c | 114 +++++++++++++++++++++--------------------------
>>> 3 files changed, 188 insertions(+), 71 deletions(-)
>>>
>>> diff --git a/fs/eventfd.c b/fs/eventfd.c
>>> index 72f5f8d..505d5de 100644
>>> --- a/fs/eventfd.c
>>> +++ b/fs/eventfd.c
>>> @@ -30,8 +30,47 @@ struct eventfd_ctx {
>>> */
>>> __u64 count;
>>> unsigned int flags;
>>> + struct srcu_struct srcu;
>>> + struct list_head nh;
>>> + struct eventfd_notifier notifier;
>>> };
>>>
>>> +static void _eventfd_wqh_notify(struct eventfd_notifier *en)
>>> +{
>>> + struct eventfd_ctx *ctx = container_of(en,
>>> + struct eventfd_ctx,
>>> + notifier);
>>> +
>>> + if (waitqueue_active(&ctx->wqh))
>>> + wake_up_poll(&ctx->wqh, POLLIN);
>>> +}
>>> +
>>> +static void _eventfd_notify(struct eventfd_ctx *ctx)
>>> +{
>>> + struct eventfd_notifier *en;
>>> + int idx;
>>> +
>>> + idx = srcu_read_lock(&ctx->srcu);
>>> +
>>> + /*
>>> + * The goal here is to allow the notification to be preemptible
>>> + * as often as possible. We cannot achieve this with the basic
>>> + * wqh mechanism because it requires the wqh->lock. Therefore
>>> + * we have an internal srcu list mechanism of which the wqh is
>>> + * a client.
>>> + *
>>> + * Not all paths will invoke this function in process context.
>>> + * Callers should check for suitable state before assuming they
>>> + * can sleep (such as with preemptible()). Paul McKenney assures
>>> + * me that srcu_read_lock is compatible with in-atomic, as long as
>>> + * the code within the critical section is also compatible.
>>> + */
>>> + list_for_each_entry_rcu(en, &ctx->nh, list)
>>> + en->ops->signal(en);
>>> +
>>> + srcu_read_unlock(&ctx->srcu, idx);
>>> +}
>>> +
>>> /*
>>> * Adds "n" to the eventfd counter "count". Returns "n" in case of
>>> * success, or a value lower then "n" in case of coutner overflow.
>>>
>>>
>> This is ugly, isn't it? With CONFIG_PREEMPT=no preemptible() is always false.
>>
>>
>
> As an aside, this is something I would like to address. I keep running
> into this pattern where I could do something in-line if I had a
> "can_sleep()" context. Otherwise, I have to punt to something like a
> workqueue which adds latency. The closest thing I have to "can_sleep()"
> is preemptible(), which, as you correctly pointed out is limited to only
> working with CONFIG_PREEMPT=y.
>
> Its been a while since I looked into it, but one of the barriers that
> would need to be overcome is the fact that the preempt_count stuff gets
> compiled away with CONFIG_PREEMPT=n. It is conceivable that we could
> make the preempt_count logic an independent config variable from
> CONFIG_PREEMPT to provide a can_sleep() macro without requiring
> full-blow preemption to be enabled. So my questions would be as follows:
>
> a) Is the community conducive to such an idea?
> b) Are there other things to consider/fix besides the lack of
> preempt_count in order to implement such a beast?
>
> -Greg
>
>
>
Hi Ingo,

Any thoughts here?

-Greg

Attachment: signature.asc
Description: OpenPGP digital signature