From mboxrd@z Thu Jan 1 00:00:00 1970 From: Arnd Bergmann Subject: Re: build_skb() and data corruption Date: Tue, 14 Jan 2014 16:51:34 +0100 Message-ID: <201401141651.34633.arnd@arndb.de> References: <1389620539.3720.157.camel@deadeye.wl.decadent.org.uk> Mime-Version: 1.0 Content-Type: Text/Plain; charset="utf-8" Content-Transfer-Encoding: 7bit Cc: netdev , "linux-arm-kernel@lists.infradead.org" , "linux-kernel@vger.kernel.org" , alexander.h.duyck@intel.com, Florian Fainelli , Ben Hutchings To: Jonas Jensen Return-path: In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On Tuesday 14 January 2014, Jonas Jensen wrote: > Thanks for the replies, you led me to a new solution, > > > I now think build_skb() is not the right choice, my motivation for > using it in the first place, that I thought it meant getting away with > not copying memory. > > build_skb() is replaced by netdev_alloc_skb_ip_align() and memcpy() > (derived from drivers/net/ethernet/realtek/r8169.c). > > Read errors are gone, even without syncing DMA. Is it a good idea to > do it anyway, i.e. leave calls to dma_sync_single_* in? The calls to dma_sync_single_* in the moxart_rx() function are needed. The call to arm_dma_ops.sync_single_for_device() in moxart_mac_setup_desc_ring() is wrong, because the buffer is already owned by the device at that point (just after dma_map_single), and because you should use the official dma_* api rather than using the arm_dma_ops struct. Arnd