From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-1.1 required=3.0 tests=DKIM_SIGNED,DKIM_VALID, DKIM_VALID_AU,HEADER_FROM_DIFFERENT_DOMAINS,MAILING_LIST_MULTI,SPF_PASS autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id A1223C46464 for ; Thu, 9 Aug 2018 10:27:25 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id 4087E21C16 for ; Thu, 9 Aug 2018 10:27:25 +0000 (UTC) Authentication-Results: mail.kernel.org; dkim=pass (1024-bit key) header.d=ideasonboard.com header.i=@ideasonboard.com header.b="cU3o1Cte" DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 4087E21C16 Authentication-Results: mail.kernel.org; dmarc=none (p=none dis=none) header.from=ideasonboard.com Authentication-Results: mail.kernel.org; spf=none smtp.mailfrom=linux-kernel-owner@vger.kernel.org Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1730164AbeHIMvf (ORCPT ); Thu, 9 Aug 2018 08:51:35 -0400 Received: from perceval.ideasonboard.com ([213.167.242.64]:32850 "EHLO perceval.ideasonboard.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1727858AbeHIMvf (ORCPT ); Thu, 9 Aug 2018 08:51:35 -0400 Received: from avalon.localnet (dfj612ybrt5fhg77mgycy-3.rev.dnainternet.fi [IPv6:2001:14ba:21f5:5b00:2e86:4862:ef6a:2804]) by perceval.ideasonboard.com (Postfix) with ESMTPSA id 4303C1267; Thu, 9 Aug 2018 12:27:19 +0200 (CEST) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/simple; d=ideasonboard.com; s=mail; t=1533810440; bh=40WS+82DmrRhMazPOiOxuKy+bQO4JfTnEyfuSZvHmtY=; h=From:To:Cc:Subject:Date:In-Reply-To:References:From; b=cU3o1CteUWKrSWmcnvRYCbQwfjcIbvNun7He62fP4ixmN7RUplTXC/UFwaDQlGF3c Y1hw8fZ3VqMAgIgOyAHvJ1AvBySHIFH7h+KdjdQWZuTcnsl94hov0gSLRSHqy+Ly4k xORcVIuAQLeIJ9P953DQGTaAnM/ZCDeou4L25jNU= From: Laurent Pinchart To: Tomasz Figa Cc: "Matwey V. Kornilov" , Alan Stern , Ezequiel Garcia , hdegoede@redhat.com, Hans Verkuil , Mauro Carvalho Chehab , rostedt@goodmis.org, mingo@redhat.com, Mike Isely , Bhumika Goyal , Colin King , Linux Media Mailing List , Linux Kernel Mailing List , Kieran Bingham , keiichiw@chromium.org Subject: Re: [PATCH 2/2] media: usb: pwc: Don't use coherent DMA buffers for ISO transfer Date: Thu, 09 Aug 2018 13:28:04 +0300 Message-ID: <3036308.kSDdPnDKq4@avalon> Organization: Ideas on Board Oy In-Reply-To: References: <1913405.2MshdJEm1G@avalon> MIME-Version: 1.0 Content-Transfer-Encoding: 7Bit Content-Type: text/plain; charset="us-ascii" Sender: linux-kernel-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Hi Tomasz, On Thursday, 9 August 2018 05:36:46 EEST Tomasz Figa wrote: > On Thu, Aug 9, 2018 at 7:31 AM Laurent Pinchart wrote: > > On Saturday, 4 August 2018 11:00:05 EEST Matwey V. Kornilov wrote: > >> 2018-07-30 18:35 GMT+03:00 Laurent Pinchart: > >>> On Tuesday, 24 July 2018 21:56:09 EEST Matwey V. Kornilov wrote: > >>>> 2018-07-23 21:57 GMT+03:00 Alan Stern: > >>>>> On Mon, 23 Jul 2018, Matwey V. Kornilov wrote: > >>>>>> I've tried to strategies: > >>>>>> > >>>>>> 1) Use dma_unmap and dma_map inside the handler (I suppose this is > >>>>>> similar to how USB core does when there is no > >>>>>> URB_NO_TRANSFER_DMA_MAP) > >>>>> > >>>>> Yes. > >>>>> > >>>>>> 2) Use sync_cpu and sync_device inside the handler (and dma_map > >>>>>> only once at memory allocation) > >>>>>> > >>>>>> It is interesting that dma_unmap/dma_map pair leads to the lower > >>>>>> overhead (+1us) than sync_cpu/sync_device (+2us) at x86_64 > >>>>>> platform. At armv7l platform using dma_unmap/dma_map leads to ~50 > >>>>>> usec in the handler, and sync_cpu/sync_device - ~65 usec. > >>>>>> > >>>>>> However, I am not sure is it mandatory to call > >>>>>> dma_sync_single_for_device for FROM_DEVICE direction? > >>>>> > >>>>> According to Documentation/DMA-API-HOWTO.txt, the CPU should not > >>>>> write to a DMA_FROM_DEVICE-mapped area, so > >>>>> dma_sync_single_for_device() is not needed. > >>>> > >>>> Well, I measured the following at armv7l. The handler execution time > >>>> (URB_NO_TRANSFER_DMA_MAP is used for all cases): > >>>> > >>>> 1) coherent DMA: ~3000 usec (pwc is not functional) > >>>> 2) explicit dma_unmap and dma_map in the handler: ~52 usec > >>>> 3) explicit dma_sync_single_for_cpu (no dma_sync_single_for_device): > >>>> ~56 usec > >>> > >>> I really don't understand why the sync option is slower. Could you > >>> please investigate ? Before doing anything we need to make sure we have > >>> a full understanding of the problem. > >> > >> Hi, > >> > >> I've found one drawback in my measurements. I forgot to fix CPU > >> frequency at lowest state 300MHz. Now, I remeasured > >> > >> 2) dma_unmap and dma_map in the handler: > >> 2A) dma_unmap_single call: 28.8 +- 1.5 usec > >> 2B) memcpy and the rest: 58 +- 6 usec > >> 2C) dma_map_single call: 22 +- 2 usec > >> Total: 110 +- 7 usec > >> > >> 3) dma_sync_single_for_cpu > >> 3A) dma_sync_single_for_cpu call: 29.4 +- 1.7 usec > >> 3B) memcpy and the rest: 59 +- 6 usec > >> 3C) noop (trace events overhead): 5 +- 2 usec > >> Total: 93 +- 7 usec > >> > >> So, now we see that 2A and 3A (as well as 2B and 3B) agree good within > >> error ranges. > > > > Thank you for the time you've spent on these measurements, the information > > is useful and your work very appreciated. > > > >>>> So, I suppose that unfortunately Tomasz suggestion doesn't work. > >>>> There is no performance improvement when dma_sync_single is used. > >>>> > >>>> At x86_64 the following happens: > >>>> > >>>> 1) coherent DMA: ~2 usec > >>> > >>> What do you mean by coherent DMA for x86_64 ? Is that > >>> usb_alloc_coherent() ? Could you trace it to see how memory is allocated > >>> exactly, and how it's mapped to the CPU ? I suspect that it will end up > >>> in dma_direct_alloc() but I'd like a confirmation. > >> > >> usb_alloc_coherents() ends up inside hcd_buffer_alloc() where > >> dma_alloc_coherent() is called. Keep in mind, that requested size is > >> 9560 in our case and pool is not used. > >> > >>>> 2) explicit dma_unmap and dma_map in the handler: ~3.5 usec > >>>> 3) explicit dma_sync_single_for_cpu (no dma_sync_single_for_device): > >>>> ~4 usec > >>>> > >>>> So, whats to do next? Personally, I think that DMA streaming API > >>>> introduces not so great overhead. > >>> > >>> It might not be very large, but with USB3 cameras at high resolutions > >>> and framerates, it might still become noticeable. I wouldn't degrade > >>> performances on x86, especially if we can decide which option to use > >>> based on the platform (or perhaps even better based on Kconfig options > >>> such as DMA_NONCOHERENT). > >> > >> PWC is discontinued chip, so there will not be any new USB3 cameras. > > > > You're right. I had in mind other USB cameras that would benefit from the > > same change, and in particular the uvcvideo driver, which is used by USB3 > > cameras. > > > >> Kconfig won't work here, as I said before, DMA config is stored inside > >> device tree blob on ARM architecture. > > > > But couldn't we skip it at least on x86 ? > > If we use the map-once, sync-repeatedly approach, would there be > anything to gain on x86? I believe the sync ops there would be > effectively no-ops, so the only overhead would be of a function call. With that approach, and iff they're effectively no-ops, that should be fine. We thus need to double-check. -- Regards, Laurent Pinchart