| Message ID | 1444742316-27986-1-git-send-email-laurent.pinchart@ideasonboard.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hello everybody, Ping ? On Tuesday 13 October 2015 16:18:36 Laurent Pinchart wrote: > The maximum DMA segment size controls the IOMMU mapping granularity. Its > default value is 64kB, resulting in potentially non-contiguous IOMMU > mappings. Configure it to 4GB to ensure that buffers get mapped > contiguously. > > Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> > --- > drivers/media/platform/omap3isp/isp.c | 4 ++++ > drivers/media/platform/omap3isp/isp.h | 1 + > 2 files changed, 5 insertions(+) > > I'm posting this as an RFC because I'm not happy with the patch, even if it > gets the job done. > > On ARM the maximum DMA segment size is used when creating IOMMU mappings. As > a large number of devices require contiguous memory buffers (this is a very > common requirement for video-related embedded devices) the default 64kB > value doesn't work. > > I haven't investigated the history behind this API in details but I have a > feeling something is not quite right. We force most drivers to explicitly > set the maximum segment size from a default that seems valid for specific > use cases only. Furthermore, as the DMA parameters are not stored directly > in struct device this require allocation of external memory for which we > have no proper management rule, making automatic handling of the DMA > parameters in frameworks or helper functions cumbersome (for a discussion > on this topic see http://www.spinics.net/lists/linux-media/msg92467.html > and http://lists.infradead.org/pipermail/linux-arm-kernel/2014-> November/305913.html). > > Is it time to fix this mess ? > > diff --git a/drivers/media/platform/omap3isp/isp.c > b/drivers/media/platform/omap3isp/isp.c index 17430a6ed85a..ebf7dc76e94d > 100644 > --- a/drivers/media/platform/omap3isp/isp.c > +++ b/drivers/media/platform/omap3isp/isp.c > @@ -2444,6 +2444,10 @@ static int isp_probe(struct platform_device *pdev) > if (ret) > goto error; > > + isp->dev->dma_parms = &isp->dma_parms; > + dma_set_max_seg_size(isp->dev, DMA_BIT_MASK(32)); > + dma_set_seg_boundary(isp->dev, 0xffffffff); > + > platform_set_drvdata(pdev, isp); > > /* Regulators */ > diff --git a/drivers/media/platform/omap3isp/isp.h > b/drivers/media/platform/omap3isp/isp.h index e579943175c4..4b2231cf01be > 100644 > --- a/drivers/media/platform/omap3isp/isp.h > +++ b/drivers/media/platform/omap3isp/isp.h > @@ -193,6 +193,7 @@ struct isp_device { > u32 syscon_offset; > u32 phy_type; > > + struct device_dma_parameters dma_parms; > struct dma_iommu_mapping *mapping; > > /* ISP Obj */
Hi Laurent, On 09/11/15 14:18, Laurent Pinchart wrote: > Hello everybody, > > Ping ? Apologies, I did start writing a response a while ago, but it ended up getting subsumed into the bigger DMA API discussion thread. > On Tuesday 13 October 2015 16:18:36 Laurent Pinchart wrote: >> The maximum DMA segment size controls the IOMMU mapping granularity. Its >> default value is 64kB, resulting in potentially non-contiguous IOMMU >> mappings. Configure it to 4GB to ensure that buffers get mapped >> contiguously. >> >> Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> >> --- >> drivers/media/platform/omap3isp/isp.c | 4 ++++ >> drivers/media/platform/omap3isp/isp.h | 1 + >> 2 files changed, 5 insertions(+) >> >> I'm posting this as an RFC because I'm not happy with the patch, even if it >> gets the job done. >> >> On ARM the maximum DMA segment size is used when creating IOMMU mappings. As >> a large number of devices require contiguous memory buffers (this is a very >> common requirement for video-related embedded devices) the default 64kB >> value doesn't work. Per the initial patch (6b7b65105522), dma_parms was intended to expose hardware limitations of scatter-gather-capable devices, to prevent DMA API implementations from merging segments beyond a device's limits. Thus the way 32-bit ARM (and seemingly noone else) is taking it as something to apply to non-scatter-gather-capable devices in order to force the DMA API implementation to merge segments seems very questionable. There's nothing in the streaming DMA API which gives any guarantee of a contiguous mapping, so it's the incorrect assumption that it does which needs fixing. Whether we rework the callers or the API itself is the open question there, I think. >> I haven't investigated the history behind this API in details but I have a >> feeling something is not quite right. We force most drivers to explicitly >> set the maximum segment size from a default that seems valid for specific >> use cases only. Furthermore, as the DMA parameters are not stored directly >> in struct device this require allocation of external memory for which we >> have no proper management rule, making automatic handling of the DMA >> parameters in frameworks or helper functions cumbersome (for a discussion >> on this topic see http://www.spinics.net/lists/linux-media/msg92467.html >> and http://lists.infradead.org/pipermail/linux-arm-kernel/2014-> November/305913.html). >> >> Is it time to fix this mess ? I agree that it would certainly be preferable to tackle the underlying issue instead of adding more point hacks to further entrench non-portable code into the kernel. In terms of modifying the API, the most reasonable idea which comes to mind would be a DMA attribute, and on closer inspection, I see that DMA_ATTR_FORCE_CONTIGUOUS is already a thing - perhaps we should weigh up whether coherent and streaming DMA could overload the same attribute with subtly different meanings, or whether we'd want e.g. DMA_ATTR_FORCE_PHYS_CONTIGUOUS and DMA_ATTR_FORCE_DMA_CONTIGUOUS to coexist. Robin. >> diff --git a/drivers/media/platform/omap3isp/isp.c >> b/drivers/media/platform/omap3isp/isp.c index 17430a6ed85a..ebf7dc76e94d >> 100644 >> --- a/drivers/media/platform/omap3isp/isp.c >> +++ b/drivers/media/platform/omap3isp/isp.c >> @@ -2444,6 +2444,10 @@ static int isp_probe(struct platform_device *pdev) >> if (ret) >> goto error; >> >> + isp->dev->dma_parms = &isp->dma_parms; >> + dma_set_max_seg_size(isp->dev, DMA_BIT_MASK(32)); >> + dma_set_seg_boundary(isp->dev, 0xffffffff); Whatever happens, 002edb6f6f2a should now make this line redundant :D >> + >> platform_set_drvdata(pdev, isp); >> >> /* Regulators */ >> diff --git a/drivers/media/platform/omap3isp/isp.h >> b/drivers/media/platform/omap3isp/isp.h index e579943175c4..4b2231cf01be >> 100644 >> --- a/drivers/media/platform/omap3isp/isp.h >> +++ b/drivers/media/platform/omap3isp/isp.h >> @@ -193,6 +193,7 @@ struct isp_device { >> u32 syscon_offset; >> u32 phy_type; >> >> + struct device_dma_parameters dma_parms; >> struct dma_iommu_mapping *mapping; >> >> /* ISP Obj */ >
Hi Laurent, I really have no idea how this problem should be addressed. I've already proposed to handle it in videobuf2-dc, but this has been rejected. Maybe the only way will be to add dma_set_max_seg_size(isp->dev, DMA_BIT_MASK(32)) to every v4l2 driver, which use videobuf2-dc and add a check in videobuf2-dc. Best regards
diff --git a/drivers/media/platform/omap3isp/isp.c b/drivers/media/platform/omap3isp/isp.c index 17430a6ed85a..ebf7dc76e94d 100644 --- a/drivers/media/platform/omap3isp/isp.c +++ b/drivers/media/platform/omap3isp/isp.c @@ -2444,6 +2444,10 @@ static int isp_probe(struct platform_device *pdev) if (ret) goto error; + isp->dev->dma_parms = &isp->dma_parms; + dma_set_max_seg_size(isp->dev, DMA_BIT_MASK(32)); + dma_set_seg_boundary(isp->dev, 0xffffffff); + platform_set_drvdata(pdev, isp); /* Regulators */ diff --git a/drivers/media/platform/omap3isp/isp.h b/drivers/media/platform/omap3isp/isp.h index e579943175c4..4b2231cf01be 100644 --- a/drivers/media/platform/omap3isp/isp.h +++ b/drivers/media/platform/omap3isp/isp.h @@ -193,6 +193,7 @@ struct isp_device { u32 syscon_offset; u32 phy_type; + struct device_dma_parameters dma_parms; struct dma_iommu_mapping *mapping; /* ISP Obj */
The maximum DMA segment size controls the IOMMU mapping granularity. Its default value is 64kB, resulting in potentially non-contiguous IOMMU mappings. Configure it to 4GB to ensure that buffers get mapped contiguously. Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com> --- drivers/media/platform/omap3isp/isp.c | 4 ++++ drivers/media/platform/omap3isp/isp.h | 1 + 2 files changed, 5 insertions(+) I'm posting this as an RFC because I'm not happy with the patch, even if it gets the job done. On ARM the maximum DMA segment size is used when creating IOMMU mappings. As a large number of devices require contiguous memory buffers (this is a very common requirement for video-related embedded devices) the default 64kB value doesn't work. I haven't investigated the history behind this API in details but I have a feeling something is not quite right. We force most drivers to explicitly set the maximum segment size from a default that seems valid for specific use cases only. Furthermore, as the DMA parameters are not stored directly in struct device this require allocation of external memory for which we have no proper management rule, making automatic handling of the DMA parameters in frameworks or helper functions cumbersome (for a discussion on this topic see http://www.spinics.net/lists/linux-media/msg92467.html and http://lists.infradead.org/pipermail/linux-arm-kernel/2014-November/305913.html). Is it time to fix this mess ?