From patchwork Sun Sep 20 11:23:11 2020 Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: 7bit X-Patchwork-Submitter: Serge Semin X-Patchwork-Id: 11787559 Return-Path: Received: from mail.kernel.org (pdx-korg-mail-1.web.codeaurora.org [172.30.200.123]) by pdx-korg-patchwork-2.web.codeaurora.org (Postfix) with ESMTP id BF972139F for ; Sun, 20 Sep 2020 11:32:29 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [23.128.96.18]) by mail.kernel.org (Postfix) with ESMTP id AD00120EDD for ; Sun, 20 Sep 2020 11:32:29 +0000 (UTC) Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1726397AbgITLc3 (ORCPT ); Sun, 20 Sep 2020 07:32:29 -0400 Received: from mail.baikalelectronics.com ([87.245.175.226]:54050 "EHLO mail.baikalelectronics.ru" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1726314AbgITLc2 (ORCPT ); Sun, 20 Sep 2020 07:32:28 -0400 Received: from localhost (unknown [127.0.0.1]) by mail.baikalelectronics.ru (Postfix) with ESMTP id 8F8DF80307CB; Sun, 20 Sep 2020 11:23:26 +0000 (UTC) X-Virus-Scanned: amavisd-new at baikalelectronics.ru Received: from mail.baikalelectronics.ru ([127.0.0.1]) by localhost (mail.baikalelectronics.ru [127.0.0.1]) (amavisd-new, port 10024) with ESMTP id wGiTZSOb0tRl; Sun, 20 Sep 2020 14:23:25 +0300 (MSK) From: Serge Semin To: Mark Brown CC: Serge Semin , Serge Semin , Alexey Malahov , Georgy Vlasov , Ramil Zaripov , Pavel Parkhomenko , Peter Ujfalusi , Andy Shevchenko , Andy Shevchenko , Feng Tang , Vinod Koul , , Subject: [PATCH v2 00/11] spi: dw-dma: Add max SG entries burst capability support Date: Sun, 20 Sep 2020 14:23:11 +0300 Message-ID: <20200920112322.24585-1-Sergey.Semin@baikalelectronics.ru> MIME-Version: 1.0 X-ClientProxiedBy: MAIL.baikal.int (192.168.51.25) To mail (192.168.51.25) Precedence: bulk List-ID: X-Mailing-List: linux-spi@vger.kernel.org Mainly this series is about fixing a very nasty problem discovered in the DW APB SSI driver working in a couple with DW DMAC, which doesn't have multi-block capability support (a.k.a. accelerated SG list entries traversal/burst, or automatic LLP entries reload, etc.). DW DMA IP-core and its DMA channel on the synthesize stage can be tuned by setting a vast number of the model parameters, some of which are provided to create a more optimized/performant controller. In particular two of those parameters are DMAH_CHx_MULTI_BLK_EN and DMAH_CHx_HC_LLP. If at least one of them is set to zero (false) then the target DMA controller will be smaller and faster but will lack of the DMA blocks chaining support. In the kernel notation it means that the controller won't be able to automatically reload a next SG-list entry, when a current one is finished. Since Linux kernel DMA subsystem interface requires to have the SG-lists execution supported, the DW DMA engine driver is developed in a way so to resubmit each SG-list entry one-by-one by software means: each SG-list entry execution finish is tracked by the DW DMA controller interrupt, which handler executes a tasklet, which then re-charges a DW DMA channel with a next SG-list entry if one is available. Such implementation is a normal design and works well for the most of the DW DMAC client devices. But it may cause problems for devices, which send and receive data by means of internal FIFOs. That requires having several DMA channels working synchronously in order to prevent the FIFOs overflow. A bright example of such device is the DW APB SSI controller. It has Tx and Rx FIFOs, which first need to be filled in with data before data sending out or receiving in. But those FIFOs are inter-dependent because of the SPI nature and its shift-register design. So each sent over Tx FIFO byte immediately causes getting a byte from the SPI bus into the Rx FIFO. It makes a strategy of working with the SPI controller a bit tricky. The more data we push into the Tx FIFO and the faster the SPI bus is, the more careful we have to be in pulling data from Rx FIFO since if software or DMA engine misses a moment when the Rx FIFO is full, for instance, due to being busy with some other activity or due to being blocked if system bus is busy with doing something else or just due to being too slow to keep up with incoming data, then Rx FIFO will be overflown, which consequently causes data loss. Needless to say that such situation is fatal and mustn't be tolerated for a bus like SPI. In application to the DW APB SSI driver the problem above may happen when DW DMAC is synthesized with no multi-block capability support and it's enabled to be working with DW APB SSI for full-duplex transfers. DW APB SSI driver allocates two DW DMAC channels to perform Tx and Rx SPI transfers, initializes them, submits Tx and Rx SG-lists for execution and then waits until the DMA transfers are finished. The issue happens when Rx SG-list consists of more than one entry. Since the multi-block capability is unsupported the DW DMAC driver will use the software-based SG-list entries traverse implementation, which by design may cause non-deterministic latency during the Rx SG-list entries re-charge. During the entries re-charge procedure the DMA Tx channel will keep pushing data into the SPI Tx FIFO. DW APB SSI controller in its turn will keep pushing data out from the Tx FIFO to the SPI bus, and will immediately fill in the Rx FIFO. So if the latency is big enough, then we'll eventually end up with Rx FIFO overflow. One of the possible solution of the problem denoted above is to feed the DMA engine with the Tx and Rx SG-list entries one-by-one. This patch-series provides an implementation of that approach. First it moves the requested DMA channels configuration into the dma_setup() callback, which should have been there in the first place. Then it's better to move the DMA transfers submission into the DMA-preparation methods to collect all the setups in a single method. After that the current implementation of a straightforward SG-lists DMA transfer should be unpinned into a dedicated method dw_spi_dma_transfer_all() since we are about to introduce an alternative DMA-based SPI transfer approach. Since DMA-transfers finish is now locally detected we can simplify the driver code a bit and move the DMAC register cleanup to a single place in the dw_spi_dma_transfer_all() method. In order to re-use the DMA-{wait, submit Tx, submit Rx} methods we have to alter their prototypes, so they would accept SG-lists instead of the SPI-transfer structure. Finally we introduce a new DMA-based SPI-transfer method, which traverses the SG-list entries in a loop and synchronously submits each of then to the Tx and Rx DMA channels in a way so the DMA engine wouldn't need to activate the prone to errors in our case SG-list entries re-charge implementation. That new method is utilized only for the DMA controllers, which can't handle all Tx and Rx SPI transfer SG-lists in a single DMA transaction without software intervention, and for the full-duplex SPI-transfers. Link: https://lore.kernel.org/linux-spi/20200731075953.14416-1-Sergey.Semin@baikalelectronics.ru Changelog v2: - Replace negative conditional statements with the positive ones in the dw_spi_dma_submit_{tx,rx}() methods. - Terminate the prepared DMA Tx-descriptors on submission errors. - Split the patch "spi: dw-dma: Move DMA transfers submission to the channels prep methods" up into a series of more simple commits. Signed-off-by: Serge Semin Cc: Alexey Malahov Cc: Georgy Vlasov Cc: Ramil Zaripov Cc: Pavel Parkhomenko Cc: Peter Ujfalusi Cc: Andy Shevchenko Cc: Andy Shevchenko Cc: Feng Tang Cc: Vinod Koul Cc: linux-spi@vger.kernel.org Cc: linux-kernel@vger.kernel.org Serge Semin (11): spi: dw-dma: Set DMA Level registers on init spi: dw-dma: Fail DMA-based transfer if no Tx-buffer specified spi: dw-dma: Configure the DMA channels in dma_setup spi: dw-dma: Check rx_buf availability in the xfer method spi: dw-dma: Move DMA transfers submission to the channels prep methods spi: dw-dma: Check DMA Tx-desc submission status spi: dw-dma: Remove DMA Tx-desc passing around spi: dw-dma: Detach DMA transfer into a dedicated method spi: dw-dma: Move DMAC register cleanup to DMA transfer method spi: dw-dma: Pass exact data to the DMA submit and wait methods spi: dw-dma: Add one-by-one SG list entries transfer drivers/spi/spi-dw-dma.c | 316 ++++++++++++++++++++++++++++++--------- drivers/spi/spi-dw.h | 1 + 2 files changed, 245 insertions(+), 72 deletions(-)