Re: [dpdk-dev] [PATCH v6 2/2] eal: add asynchronous request API to DPDK IPC

[Thomas Monjalon <thomas@monjalon.net>]

28/03/2018 11:21, Burakov, Anatoly:
> On 28-Mar-18 9:55 AM, Tan, Jianfeng wrote:
> > On 3/28/2018 4:22 PM, Van Haaren, Harry wrote:
> >> From: Thomas Monjalon [mailto:thomas@monjalon.net]
> >>> 28/03/2018 04:08, Tan, Jianfeng:
> >>>> From: Thomas Monjalon [mailto:thomas@monjalon.net]
> >>>>> 27/03/2018 15:59, Anatoly Burakov:
> >>>>>> Under the hood, we create a separate thread to deal with replies to
> >>>>>> asynchronous requests, that will just wait to be notified by the
> >>>>>> main thread, or woken up on a timer.
> >>>>> 
> >>>>> I really don't like that a library is creating a thread.
> >>>>> We don't even know where the thread is created (which core).
> >>>>> Can it be a rte_service? or in the interrupt thread?
> >>>> 
> >>>> Agree that we'd better not adding so many threads in a library.
> >>>>
> >>>> I was considering to merge all the threads into the interrupt thread,
> >>> however, we don't have an interrupt thread in freebsd. Further, we don't
> >>> implement alarm API in freebsd. That's why I tend to current 
> >>> implementation,
> >>> and optimize it later.
> >>>
> >>> I would prefer we improve the current code now instead of polluting more
> >>> with more uncontrolled threads.
> >>>
> >>>> For rte_service, it may be not a good idea to reply on it as it needs
> >>> explicit API calls to setup.
> >>>
> >>> I don't see the issue of the explicit API.
> >>> The IPC is a new service.
> > 
> > My concern is that not every DPDK application sets up rte_service, but 
> > IPC will be used for very fundamental functions, like memory allocation. 
> > We could not possibly ask all DPDK applications to add rte_service now.
> > 
> > And also take Harry's comments below into consideration, most likely, we 
> > will move these threads into interrupt thread now by adding
> > 
> >> Although I do like to see new services, if we want to enable "core" 
> >> dpdk functionality with Services, we need a proper designed solution 
> >> for that. Service cores is not intended for "occasional" work - there 
> >> is no method to block and sleep on a specific service until work 
> >> becomes available, so this would imply a busy-polling. Using a service 
> >> (hence busy polling) for rte_malloc()-based memory mapping requests is 
> >> inefficient, and total overkill :)
> >>
> >> For this patch I suggest to use some blocking-read capable mechanism.
> > 
> > The problem here is that we add too many threads; blocking-read does not 
> > decrease # of threads.
> > 
> >>
> >> The above said, in the longer term it would be good to have a design 
> >> that allows new file-descriptors to be added to a "dpdk core" thread, 
> >> which performs occasional lengthy work if the FD has data available.
> > 
> > Interrupt thread vs rte_service, which is the direction to go? We 
> > actually have some others threads, in vhost and even virtio-user; we can 
> > also avoid those threads if we have a clear direction.
> > 
> > Thanks,
> > Jianfeng
> > 
> 
> Hi all,
> 
> First of all, @Thomas, this is not a "new library" - it's part of EAL.

I did not say it is a new library.

> We're going to be removing a few threads from EAL as it is because of 
> IPC (Jianfeng has already submitted patches for those),

I don't understand.
Which threads are you going to remove? Which patch?

> so i don't think 
> it's such a big deal to have two IPC threads instead of one. I'm open to 
> suggestions on how to make this work without a second thread, but i 
> don't see it.

I am not against the second thread.
I am against both threads :)

> We've discussed possibility of using rte_service internally, but decided 
> against it for reasons already outlined by Harry - it's not a suitable 
> mechanism for this kind of thing, not as it is.
> 
> Using interrupt thread for this _will_ work, however this will require a 
> a lot more changes, as currently alarm API allocates everything through 
> rte_malloc, while we want to use IPC for rte_malloc (which would make it 
> a circular dependency). So it'll probably be more API and more 
> complexity for dealing with malloc vs rte_malloc allocations. Hence the 
> least-bad approach taken here: a new thread.

If everybody is happy enough with "least bad" design and not trying
to improve the core design, what can I say?