MIME-Version: 1.0
In-Reply-To: <9a809c6e-3b01-4deb-e69e-65aa62d1e5e1@arm.com>
References: <1488538070-12549-1-git-send-email-sunil.kovvuri@gmail.com>
 <1488538070-12549-2-git-send-email-sunil.kovvuri@gmail.com>
 <20170303.095635.355325147613813698.davem@davemloft.net> <CA+sq2CcFxk=LuqYhR-sFaVVSu3rMaZUcytx-3tC9kEiRFNwREg@mail.gmail.com>
 <cad7da5d-fce7-cef2-d0db-9585c31f0575@arm.com> <CA+sq2CeiXZNwVsURxNNE+vxf3uhjM=vnKtb2nhtLRj6QwHKWrw@mail.gmail.com>
 <9a809c6e-3b01-4deb-e69e-65aa62d1e5e1@arm.com>
From: Sunil Kovvuri <sunil.kovvuri@gmail.com>
Date: Tue, 7 Mar 2017 18:14:31 +0530
Message-ID: <CA+sq2CfCS1aQGtCwVxhzMU3QmKy+tE6j2Zvq=U9aTodvDTFdvA@mail.gmail.com>
Subject: Re: [PATCH 1/4] net: thunderx: Fix IOMMU translation faults
To: Robin Murphy <robin.murphy@arm.com>
Cc: David Miller <davem@davemloft.net>,
        Linux Netdev List <netdev@vger.kernel.org>,
        Sunil Goutham <sgoutham@cavium.com>,
        LKML <linux-kernel@vger.kernel.org>,
        LAKML <linux-arm-kernel@lists.infradead.org>
Content-Type: text/plain; charset=UTF-8
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1685
Lines: 37

On Mon, Mar 6, 2017 at 10:02 PM, Robin Murphy <robin.murphy@arm.com> wrote:
> On 06/03/17 12:57, Sunil Kovvuri wrote:
>>>>
>>>> We are seeing a 0.75Mpps drop with IP forwarding rate due to that.
>>>> Hence I have restricted calling DMA interfaces to only when IOMMU is enabled.
>>>
>>> What's 0.07Mpps as a percentage of baseline? On a correctly configured
>>> coherent arm64 system, in the absence of an IOMMU, dma_map_*() is
>>> essentially just virt_to_phys() behind a function call or two, so I'd be
>>> interested to know where any non-trivial overhead might be coming from.
>>
>> It's a 5% drop and yes device is configured as coherent.
>> And the drop is due to additional function calls.
>
> OK, interesting - sounds like there's potential for some optimisation
> there as well. AFAICS the callchain goes:
>
> dma_map_single_attrs (inline)
> - ops->map_page (__swiotlb_map_page)
>   - swiotlb_map_page
>     - phys_to_dma (inline)
>     - dma_capable (inline)
>
> Do you happen to have a breakdown of where the time goes? If it's mostly
> just in the indirect branch our options are limited (I'm guessing
> ThunderX doesn't have a particularly fancy branch predictor, if it's not
> even got a data prefetcher), but if it's in the SWIOTLB code then
> there's certainly room for improvement (which will hopefully tie in with
> some DMA ops work I'm planning to do soon anyway).

It's the branching which is costing the performance, as you said nothing
much can be done in the common code for this. Anyway I have submitted
new patch without conditional calling of DMA APIs, will look into reducing
performance impact (if possible implement recycling) a bit later.

Thanks,
Sunil.