On 4/19/2016 2:22 PM, Christoph Hellwig wrote: > What I think we need is something like the patch below. In the long > ru nwe should also kill the mlx4_buf structure which now is pretty > pointless.
Maybe; this could be the correct approach if we can guarantee that the architecture can allocate the requested amount of memory with dma_alloc_coherent. When I brought this issue a year ago, the objection was that my code doesn't compile on intel (dma_to_phys) and also some arches run out of DMA memory with existing customer base. If there is a real need to maintain compatibility with the existing architectures due to limited amount of DMA memory, we need to correct this code to make proper use of vmap via alloc_pages and also insert the dma_sync in proper places for DMA API conformance. Also, the tx descriptors always has to be allocated from a single DMA region or the tx code needs to be corrected to support page_list. If we can live with just using dma_alloc_coherent, your solution is better. I was trying to put this support for 64bit arches only while maintaining support for the existing code base. > > --- > From a493881d2a6c90152d3daabb7b6b3afd1d254d78 Mon Sep 17 00:00:00 2001 > From: Christoph Hellwig <h...@lst.de> > Date: Tue, 19 Apr 2016 14:12:14 -0400 > Subject: mlx4_en: don't try to split and vmap dma coherent allocations > > The memory returned by dma_alloc_coherent is not suitable for calling > virt_to_page on it, as it might for example come from vmap allocator. > > Remove the code that calls virt_to_page and vmap on dma coherent > allocations from the mlx4 drivers, and replace them with a single > high-order dma_alloc_coherent call. > > Signed-off-by: Christoph Hellwig <h...@lst.de> > Reported-by: Sinan Kaya <ok...@codeaurora.org> -- Sinan Kaya Qualcomm Technologies, Inc. on behalf of Qualcomm Innovation Center, Inc. Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project
