[dpdk-dev] [PATCH v4] eal: refactor rte_eal_init into sub-functions

[Rahul Gupta]

On (01/24/24 16:53), David Marchand wrote:
> Date: Wed, 24 Jan 2024 16:53:33 +0100
> From: David Marchand <david.marchand at redhat.com>
> To: Rahul Gupta <rahulgupt at linux.microsoft.com>
> Cc: dev at dpdk.org, thomas at monjalon.net, bruce.richardson at intel.com,
>  dmitry.kozliuk at gmail.com, stephen at networkplumber.org,
>  sovaradh at linux.microsoft.com, okaya at kernel.org,
>  sujithsankar at microsoft.com, sowmini.varadhan at microsoft.com,
>  krathinavel at microsoft.com, rahulrgupta27 at gmail.com, Rahul Gupta
>  <rahulgupt at microsoft.com>
> Subject: Re: [dpdk-dev] [PATCH v4] eal: refactor rte_eal_init into
>  sub-functions
> 
> On Wed, Jan 24, 2024 at 2:45 PM Rahul Gupta
> <rahulgupt at linux.microsoft.com> wrote:
> >
> > From: Rahul Gupta <rahulgupt at microsoft.com>
> >
> > In continuation to the following email, I am sending this patch.
> > (https://inbox.dpdk.org/dev/20231110172523.GA17466@microsoft.com/)
> >
> > Initialization requires rte_eal_init + rte_pktmbuf_pool_create which
> > can consume a total time of 500-600 ms:
> > a) For many devices FLR may take a significant chunk of time
> >    (200-250 ms in our use-case), this FLR is triggered during device
> >    probe in rte_eal_init().
> > b) rte_pktmbuf_pool_create() can consume up to 300-350 ms for
> > applications that require huge memory.
> >
> > This cost is incurred on each restart (which happens in our use-case
> > during binary updates for servicing).
> > This patch provides an optimization using pthreads that applications
> > can use and which can save 200-230ms.
> >
> > In this patch, rte_eal_init() is refactored into two parts-
> > a) 1st part is dependent code ie- it’s a perquisite of the FLR and
> >    mempool creation. So this code needs to be executed before any
> >    pthreads. Its named as rte_eal_init_setup()
> > b) 2nd part of code is independent code ie- it can execute in parallel
> >    to mempool creation in a pthread. Its named as rte_eal_init_async_setup().
> >
> > In existing applications no changes are required unless they wish to leverage
> > the optimization.
> >
> > If the application wants to leverage this optimization, then it needs to call
> > rte_eal_init_async() (instead of call rte_eal_init()), then it can create a
> > thread using rte_eal_remote_launch() to schedule a task it would like todo in
> > parallel rte_eal_init_async_setup(), this task can be a mbuf pool creation
> > using- rte_pktmbuf_pool_create()
> >
> > After this, if next operations require completion of above task, then
> > user can use rte_eal_init_wait_async_setup_complete(),
> > or if user wants to just check status of that thread, then use-
> > rte_eal_init_async_setup_done()
> 
> Looking at what this patch does.. I am under the impression all you
> really need is rte_eal_init without initial probing.
> Such behavior can probably be achieved with a allowlist set to a non
> existing device (like for example "-a 0000:00:00.0"), then later, use
> device hotplug.
The patch will be useful to all the adapters irrespective of their
host plug support.
> 
> Some quick note on this patch.
> 
> - don't expose symbols externally if they are only for internal use in
> the same library,
done in next patch.
> - current version is 24.03, not 24.01 (wrt comment in version.map),
done
> - other OSes are not handled by this patch, please do the work for
> FreeBSD and Windows,
I can send patch to support non-linux OS, but due to lack of setup,
I will need help to test same.

Also, I am planning to do the porting at the end (ie after incorporating
all review comments, in order to prevent duplication of efforts).

> - as a followup of the previous point, please check if we can share
> code between OSes and, if so, move the shared code to lib/eal/common,
The code for rte_eal_init() is different for all three distros, (even if
I consider just the 1st part of rte_eal_init() ie rte_eal_init_setup()).
So its difficult to move to common dir. We will have todo it separately
for all OS.
> - did you test this series with multiprocess?
Yes it works fine, I have tested it with simple_mp app.
> - why should telemetry and other parts of the current rte_eal_init()
> be left in the second stage of this initialisation pipeline?
Actually motivation of this patch was todo most of the work in parallel
ie in second stage, so not only probe/FLR but telemetry and any other work
which can be executed in parallel are done here. (pls refer to the link
in the commit message for more details)
> 
> 
> -- 
> David Marchand


Thanks for the reviewing, my comments are inline.

Thanks,
Rahul.