Re: [PATCH v6 22/24] dmaengine: dw-edma: Bypass dma-ranges mapping for the local setup

[Serge Semin]

Robin,
Could you give your comments regarding the DMA-mask issue described
below? It would be great to resubmit a fixed revision of the series
on this week in order to make it possible to merge it into the kernel
on the next merge window.

-Sergey

On Sun, Nov 27, 2022 at 02:45:13AM +0300, Serge Semin wrote:
> On Fri, Nov 25, 2022 at 03:32:23PM +0000, Robin Murphy wrote:
> > On 2022-11-07 21:11, Serge Semin wrote:
> > > On Tue, Nov 08, 2022 at 12:04:36AM +0300, Serge Semin wrote:
> > > > DW eDMA doesn't perform any translation of the traffic generated on the
> > > > CPU/Application side. It just generates read/write AXI-bus requests with
> > > > the specified addresses. But in case if the dma-ranges DT-property is
> > > > specified for a platform device node, Linux will use it to create a
> > > > mapping the PCIe-bus regions into the CPU memory ranges. This isn't what
> > > > we want for the eDMA embedded into the locally accessed DW PCIe Root Port
> > > > and End-point. In order to work that around let's set the chan_dma_dev
> > > > flag for each DW eDMA channel thus forcing the client drivers to getting a
> > > > custom dma-ranges-less parental device for the mappings.
> > > > 
> > > > Note it will only work for the client drivers using the
> > > > dmaengine_get_dma_device() method to get the parental DMA device.
> > > 
> > > @Robin, we particularly need you opinion on this patch. I did as you
> > > said: call *_dma_configure() method to initialize the child device and
> > > set the DMA-mask here instead of the platform driver.
> > 
> 
> > Apologies, I've been busy and this series got buried in my inbox before I'd
> > clocked it as something I was supposed to be looking at.
> 
> No worries. I'm glad you responded.
> 
> > 
> > > @Vinoud, @Manivannan I had to drop your tags from this patch since its
> > > content had been significantly changed.
> > > 
> > > -Sergey
> > > 
> > > > 
> > > > Signed-off-by: Serge Semin <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx>
> > > > 
> > > > ---
> > > > 
> > > > Changelog v2:
> > > > - Fix the comment a bit to being clearer. (@Manivannan)
> > > > 
> > > > Changelog v3:
> > > > - Conditionally set dchan->dev->device.dma_coherent field since it can
> > > >    be missing on some platforms. (@Manivannan)
> > > > - Remove Manivannan' rb and tb tags since the patch content has been
> > > >    changed.
> > > > 
> > > > Changelog v6:
> > > > - Directly call *_dma_configure() method on the child device used for
> > > >    the DMA buffers mapping. (@Robin)
> > > > - Explicitly set the DMA-mask of the child device in the channel
> > > >    allocation proecedure. (@Robin)
> > > > - Drop @Manivannan and @Vinod rb- and ab-tags due to significant patch
> > > >    content change.
> > > > ---
> > > >   drivers/dma/dw-edma/dw-edma-core.c | 44 ++++++++++++++++++++++++++++++
> > > >   1 file changed, 44 insertions(+)
> > > > 
> > > > diff --git a/drivers/dma/dw-edma/dw-edma-core.c b/drivers/dma/dw-edma/dw-edma-core.c
> > > > index e3671bfbe186..846518509753 100644
> > > > --- a/drivers/dma/dw-edma/dw-edma-core.c
> > > > +++ b/drivers/dma/dw-edma/dw-edma-core.c
> > > > @@ -6,9 +6,11 @@
> > > >    * Author: Gustavo Pimentel <gustavo.pimentel@xxxxxxxxxxxx>
> > > >    */
> > > > +#include <linux/acpi.h>
> > > >   #include <linux/module.h>
> > > >   #include <linux/device.h>
> > > >   #include <linux/kernel.h>
> > > > +#include <linux/of_device.h>
> > > >   #include <linux/dmaengine.h>
> > > >   #include <linux/err.h>
> > > >   #include <linux/interrupt.h>
> > > > @@ -711,10 +713,52 @@ static irqreturn_t dw_edma_interrupt_common(int irq, void *data)
> > > >   static int dw_edma_alloc_chan_resources(struct dma_chan *dchan)
> > > >   {
> > > >   	struct dw_edma_chan *chan = dchan2dw_edma_chan(dchan);
> > > > +	struct device *dev = chan->dw->chip->dev;
> > > > +	int ret;
> > > >   	if (chan->status != EDMA_ST_IDLE)
> > > >   		return -EBUSY;
> > > > +	/* Bypass the dma-ranges based memory regions mapping for the eDMA
> > > > +	 * controlled from the CPU/Application side since in that case
> > > > +	 * the local memory address is left untranslated.
> > > > +	 */
> > > > +	if (chan->dw->chip->flags & DW_EDMA_CHIP_LOCAL) {
> 
> 
> > > > +		ret = dma_coerce_mask_and_coherent(&dchan->dev->device,
> > > > +						   DMA_BIT_MASK(64));
> > > > +		if (ret) {
> > 
> > Setting a 64-bit mask should never fail, especially on any platform that
> > will actually run this code.
> > 
> > > > +			ret = dma_coerce_mask_and_coherent(&dchan->dev->device,
> > > > +							   DMA_BIT_MASK(32));
> 
> Indeed. I can just drop the 32-bit mask test then. (But I'd retain the
> error check anyway.)
> 
> The problem is that actual device DMA-addressing capability is
> determined by the MASTER_BUS_ADDR_WIDTH IP-core synthesize parameter.
> I can't predict its value from this generic code since it isn't
> auto-detectable and is platform-specific. That's why back then in
> our discussion I was insisting on setting the mask in the low-level
> device drivers. But after the commit 423511ec23e2 ("PCI: dwc: Drop
> dependency on ZONE_DMA32") it turned to be pointless now since the
> DMA-mask would be overwritten by the generic DW PCIe driver code anyway.
> What do you suggest then in this regard? Just keep setting the 64-bit
> mask only? This will work for my platform, but will fail for the
> devices with AXI-bus address of only 32-bits width.
> 
> > > > +			if (ret)
> > > > +				return ret;
> > > > +		}
> > > > +
> > > > +		if (dev_of_node(dev)) {
> > > > +			struct device_node *node = dev_of_node(dev);
> > > > +
> > > > +			ret = of_dma_configure(&dchan->dev->device, node, true);
> > > > +		} else if (has_acpi_companion(dev)) {
> > 
> 
> > Can this can ever happen? AFAICS there's no ACPI binding to match and probe
> > the DWC driver, at best it could only probe as a standard PNP0A08 host
> > bridge which wouldn't know anything about eDMA anyway.
> 
> There are several ACPI-based platforms with DW PCIe controllers:
> pcie-tegra194-acpi.c, pcie-al.c, pcie-hisi.c. All of them are fully
> ECAM-based so no DW eDMA probing from the Linux kernel implied. But
> these are still DW PCIe controllers and they or some other ones can
> have eDMA embedded. Do you think it won't be ever possible to either
> directly handle these controllers (bypassing the ECAM interface) or
> have a DW PCIe device accessed via the ACPI bindings?
> 
> Note basically what I've implemented here was based on the
> platform_dma_configure() DMA-configuration code pattern. I thought it
> was a reasonable choice since this code path is executed for the
> platform devices only (implied by the DW_EDMA_CHIP_LOCAL flag
> semantic).
> 
> On the second thought if the problem in subject is only specific to
> the DT-based platforms, then I could just skip channel device
> initialization here for the platform devices with no OF-node detected.
> So the question is is it specific to the DT-based platforms only?
> 
> (Before answering to the question above please read the last comment
> in this message.)
> 
> > 
> > > > +			struct acpi_device *adev = to_acpi_device_node(dev->fwnode);
> > > > +
> > > > +			ret = acpi_dma_configure(&dchan->dev->device,
> > > > +						 acpi_get_dma_attr(adev));
> > > > +		} else {
> > > > +			ret = -EINVAL;
> > > > +		}
> > > > +
> > > > +		if (ret)
> > > > +			return ret;
> > > > +
> > > > +		if (dchan->dev->device.dma_range_map) {
> > > > +			kfree(dchan->dev->device.dma_range_map);
> > > > +			dchan->dev->device.dma_range_map = NULL;
> > > > +		}
> > 
> 
> > Ugh, I guess this is still here because now you're passing the channel
> > device to of_dma_configure() such that it looks like a PCI child :(
> 
> No. It's still here because I successfully missed your email in my
> work inbox so I thought you didn't fix that dma-ranges peculiarity of
> the PCIe-host nodes.(
> 
> > 
> > Can we just set "chan->dev->device.of_node = dev->of_node;" beforehand so it
> > works as expected (with f1ad5338a4d5 in place) and we don't need to be
> > messing with the dma_range_map details at all? Note that that isn't as hacky
> > as it might sound - it's a relatively well-established practice in places
> > like I2C and SPI, and in this case it seems perfectly appropriate
> > semantically as well.
> 
> Of course we can. But now, thanks to your commit f1ad5338a4d5 ("of:
> Fix "dma-ranges" handling for bus controllers"), there is no point in
> any dma-ranges hack here because the dma-ranges property is no longer
> parsed for the PCIe-host platform device. I can and will just drop the
> custom DMA-channel device initialization from the patch. The only
> issue left to solve is about setting the DMA-mask. Please see my notes
> above regarding that problem.
> 
> -Serge(y)
> 
> > 
> > (And there should be no need to bother with of_node refcounting, since the
> > lifetime of the eDMA driver is bounded by the lifetime of the PCIe driver,
> > thus the lifetime of the DMA channel devices is bounded by the lifetime of
> > the PCIe platform device, which already holds a reference from
> > of_device_alloc().)
> > 
> > Thanks,
> > Robin.
> > 
> > > > +
> > > > +		dchan->dev->chan_dma_dev = true;
> > > > +	} else {
> > > > +		dchan->dev->chan_dma_dev = false;
> > > > +	}
> > > > +
> > > >   	return 0;
> > > >   }
> > > > -- 
> > > > 2.38.0
> > > > 
> > > >