From: linux@arm.linux.org.uk (Russell King - ARM Linux) Date: Tue, 24 Nov 2009 21:01:54 +0000 Subject: [PATCH v2 06/17] ARM: dma-mapping: fix for speculative accesses In-Reply-To: References: <20091123172722.GE18142@n2100.arm.linux.org.uk> <1259081261.16239.28.camel@pc1117.cambridge.arm.com> <20091124200019.GH17481@n2100.arm.linux.org.uk> Message-ID: <20091124210154.GI17481@n2100.arm.linux.org.uk> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org On Tue, Nov 24, 2009 at 03:35:13PM -0500, Nicolas Pitre wrote: > On Tue, 24 Nov 2009, Russell King - ARM Linux wrote: > > Having looked into pushing the 3rd argument down into per-CPU code, I've > > come up against two problems: > > > > 1. the dmabounce (hacky, I wish it'd die) code wants to be able to place > > the cache in the same state as if the buffer was used for DMA - it > > does this by calling dmac_clean_range() and outer_clean_range(). > > > > 2. various MTD map drivers want the ARM private flush_ioremap_region, > > which is built on top of dmac_inv_cache. > > > > You can't reliably implement either of these two functions using the new > > dmac_(map|unmap)_range interfaces without making the callers knowledgeable > > about the implementation of the map/unmap range interfaces. > > > > This means we currently end up with five DMA operation pointers in the > > cpu_cache_fns: > > > > void (*dma_map_range)(const void *, size_t, int); > > void (*dma_unmap_range)(const void *, size_t, int); > > void (*dma_inv_range)(const void *, const void *); > > void (*dma_clean_range)(const void *, const void *); > > void (*dma_flush_range)(const void *, const void *); > > > > which is rather excessive, especially as two will no longer have anything > > to do with DMA. > > > > I'm tempted to say that for 2.6.33, we'll take the performance hit for CPUs > > older than ARMv7 of needlessly doing DMA cache maintainence on both map and > > unmap, and we'll try to sort it out for 2.6.34. > > > > The alternative is that we persist with streaming DMA being broken on ARMv7 > > for 2.6.33. > > What about simply having those two problematic cases above implemented > on top of dmac_(map|unmap)_range interfaces even if that implies inner > implementation details, and stick a big fat warning besides them until > they're properly sorted out? That should affect a much reduced set of > users compared to all pre ARMv7 users. My point was it doesn't work. Not only will you have to select which of the map/unmap functions you require, but also the 3rd argument to get the desired behaviour depending on the CPU in use.