linux-kernel - Re: [PATCH 16/19] dma-iommu: don't depend on CONFIG_DMA_DIRECT

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-ID: <20190211163936.GA29978@lst.de>
Date:   Mon, 11 Feb 2019 17:39:36 +0100
From:   Christoph Hellwig <hch@....de>
To:     Robin Murphy <robin.murphy@....com>
Cc:     Christoph Hellwig <hch@....de>, Joerg Roedel <joro@...tes.org>,
        Catalin Marinas <catalin.marinas@....com>,
        Will Deacon <will.deacon@....com>,
        Tom Lendacky <thomas.lendacky@....com>,
        iommu@...ts.linux-foundation.org,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH 16/19] dma-iommu: don't depend on
 CONFIG_DMA_DIRECT_REMAP

On Wed, Feb 06, 2019 at 11:55:49AM +0000, Robin Murphy wrote:
> On 14/01/2019 09:41, Christoph Hellwig wrote:
>> For entirely dma coherent architectures there is no good reason to ever
>> remap dma coherent allocation.
>
> Yes there is, namely assembling large buffers without the need for massive 
> CMA areas and compaction overhead under memory fragmentation. That has 
> always been a distinct concern from the DMA_DIRECT_REMAP cases; they've 
> just been able to share a fair few code paths.

Well, I guess I need to reword this - there is no _requirement_ to
remap.  And x86 has been happy to not remap so far and I see absolutely
no reason to force anyone to remap.

>>  Move all the remap and pool code under
>> CONFIG_DMA_DIRECT_REMAP ifdefs, and drop the Kconfig dependency.
>
> As far as I'm concerned that splits things the wrong way. Logically, 
> iommu_dma_alloc() should always have done its own vmap() instead of just 
> returning the bare pages array, but that was tricky to resolve with the 
> design of having the caller handle everything to do with coherency (forcing 
> the caller to unpick that mapping just to remap it yet again in the 
> noncoherent case didn't seem sensible).

I don't parse this.  In the old code base before this series
iommu_dma_alloc is a relatively low-level helper allocating and mapping
pages.  And that one should have done the remapping, and in fact does
so since ("dma-iommu: refactor page array remap helpers").  It just
happens that the function is now called iommu_dma_alloc_remap.

The new iommu_dma_alloc is the high level entry point that handles
every possible case of different allocations, including those where
we do not have a virtual mapping.