Re: [PATCH 2/5] iommu/mediatek: Add mt8173 IOMMU driver

[Robin Murphy]

On 18/03/15 11:22, Yong Wu wrote:
> Hi Tomasz,
>     Thanks very much for your review. please help check below.
> The others I will fix in the next version.
>
> Hi Robin,
>     There are some place I would like you can have a look and give me
> some suggestion.
>
> On Wed, 2015-03-11 at 19:53 +0900, Tomasz Figa wrote:
> > Hi,
> >
> > Please find next part of my comments inline.
> >
> > On Fri, Mar 6, 2015 at 7:48 PM,  <yong.wu@xxxxxxxxxxxx> wrote:
> >
> > [snip]
> >
> > > +/*
> > > + * pimudev is a global var for dma_alloc_coherent.
> > > + * It is not accepatable, we will delete it if "domain_alloc" is enabled
> >
> > It looks like we indeed need to use dma_alloc_coherent() and we don't
> > have a good way to pass the device pointer to domain_init callback.
> >
> > If you don't expect SoCs in the nearest future to have multiple M4U
> > blocks, then I guess this global variable could stay, after changing
> > the comment into an explanation why it's correct. Also it should be
> > moved to the top of the file, below #include directives, as this is
> > where usually global variables are located.
> @Robin,
>       We have merged this patch[0] in order to delete the global var, But
> it seems that your patch of "arm64:IOMMU" isn't based on it right row.
> it will build fail.

Yeah, I've not yet managed to try pulling in that series (much as I 
approve of it), partly as I know doing so is going to lean towards a 
not-insignificant rework and I'd rather avoid picking up more unmerged 
dependencies to block getting _something_ in for arm64 (which we can 
then improve).

> [0]:http://lists.linuxfoundation.org/pipermail/iommu/2015-January/011939.html
>
> > > + */
> > > +static struct device *pimudev;
> > > +
> [snip]
> > > +
> > > +static int mtk_iommu_attach_device(struct iommu_domain *domain,
> > > +                                  struct device *dev)
> > > +{
> > > +       unsigned long flags;
> > > +       struct mtk_iommu_domain *priv = domain->priv;
> > > +       struct mtk_iommu_info *piommu = priv->piommuinfo;
> > > +       struct of_phandle_args out_args = {0};
> > > +       struct device *imudev;
> > > +       unsigned int i = 0;
> > > +
> > > +       if (!piommu)
> >
> > Could you explain when this can happen?
> 	If we call arch_setup_dma_ops to create a iommu domain,
> it will enter iommu_dma_attach_device, then enter here. At that time, we
> don't add the private data to this "struct iommu_domain *".
> @Robin, Could this be improved?

Calling arch_setup_dma_ops() from the driver looks plain wrong, 
especially given that you apparently attach the IOMMU to itself - if you 
want your own domain you should use iommu_dma_create_domain(). I admit 
that still leaves you having to dance around a bit in order to tear down 
the automatic domains for now, but hopefully we'll get the core code 
sorted out sooner rather than later.

> > > +               goto imudev;
> >
> > return 0;
> >
> > > +       else
> >
> > No else needed.
> >
> > > +               imudev = piommu->dev;
> > > +
> > > +       spin_lock_irqsave(&priv->portlock, flags);
> >
> > What is protected by this spinlock?
> 	We will write a register of the local arbiter while config port. If
> some modules are in the same local arbiter, it may be overwrite. so I
> add it here.
> > > +
> > > +       while (!of_parse_phandle_with_args(dev->of_node, "iommus",
> > > +                                          "#iommu-cells", i, &out_args)) {
> > > +               if (1 == out_args.args_count) {
> >
> > Can we be sure that this is actually referring to our IOMMU?
> >
> > Maybe this should be rewritten to
> >
> > if (out_args.np != imudev->of_node)
> >          continue;
> > if (out_args.args_count != 1) {
> >          dev_err(imudev, "invalid #iommu-cells property for IOMMU %s\n",
> >
> > }
> >
> > > +                       unsigned int portid = out_args.args[0];
> > > +
> > > +                       dev_dbg(dev, "iommu add port:%d\n", portid);
> >
> > imudev should be used here instead of dev.
> >
> > > +
> > > +                       mtk_iommu_config_port(piommu, portid);
> > > +
> > > +                       if (i == 0)
> > > +                               dev->archdata.dma_ops =
> > > +                                       piommu->dev->archdata.dma_ops;
> >
> > Shouldn't this be set automatically by IOMMU or DMA mapping core?
> @Robin,
>       In the original "arm_iommu_attach_device" of arm/mm, it will call
> set_dma_ops to add iommu_ops for each iommu device.
> But iommu_dma_attach_device don't help this, so I have to add it here.
> Could this be improved?

If you implemented a simple of_xlate callback so that the core code 
handles the dma_ops as intended, I think the simplest cheat would be to 
check the client device's domain, either on attachment or when they 
start mapping/unmapping, and move them to your own domain if necessary. 
I'm putting together a v3 of the DMA mapping series, so I'll have a look 
to see if I can squeeze in a way to make that a bit less painful until 
we solve it properly.


Robin.

> > > +               }
> > > +               i++;
> > > +       }
> > > +
> > > +       spin_unlock_irqrestore(&priv->portlock, flags);
> > > +
> > > +imudev:
> > > +       return 0;
> > > +}
> > > +
> > > +static void mtk_iommu_detach_device(struct iommu_domain *domain,
> > > +                                   struct device *dev)
> > > +{
> >
> > No hardware (de)configuration or clean-up necessary?
> I will add it. Actually we design like this:If a device have attached to
> iommu domain, it won't detach from it.
> > > +}
> > > +
> [snip]
> > > +
> > > +       piommu->protect_va = devm_kmalloc(piommu->dev, MTK_PROTECT_PA_ALIGN*2,
> >
> > style: Operators like * should have space on both sides.
> >
> > > +                                         GFP_KERNEL);
> >
> > Shouldn't dma_alloc_coherent() be used for this?
>       We don't care the data in it. I think they are the same. Could you
> help tell me why dma_alloc_coherent may be better.
> > > +       if (!piommu->protect_va)
> > > +               goto protect_err;
> >
> > Please return -ENOMEM here directly, as there is nothing to clean up
> > in this case.
> [snip]
> > > +               dev_err(piommu->dev, "IRQ request %d failed\n",
> > > +                       piommu->irq);
> > > +               goto hw_err;
> > > +       }
> > > +
> > > +       iommu_set_fault_handler(domain, mtk_iommu_fault_handler, piommu);
> >
> > I don't see any other drivers doing this. Isn't this for upper layers,
> > so that they can set their own generic fault handlers?
>       I think that this function is related with the iommu domain, we
> have only one multimedia iommu domain. so I add it after the iommu
> domain are created.
> > > +
> > > +       dev_set_drvdata(piommu->dev, piommu);
> >
> > This should be set before allowing the interrupt to fire. In other
> > words, the driver should be fully set up at the time of enabling the
> > IRQ.
> >
> > > +
> > > +       return 0;
> >
> > style: Missing blank line.
> >
> > > +hw_err:
> > > +       arch_teardown_dma_ops(piommu->dev);
> > > +pte_err:
> > > +       kmem_cache_destroy(piommu->m4u_pte_kmem);
> > > +protect_err:
> > > +       dev_err(piommu->dev, "probe error\n");
> >
> > Please replace this with specific messages for all errors (in case the
> > called function doesn't already print one like kmalloc and friends).
> >
> > > +       return 0;
> >
> > Returning 0, which means success, doesn't look like a good idea for
> > signalling a failure. Please return the correct error code as received
> > from function that errors out if possible.
> >
> > End of part 3.
> >
> > Best regards,
> > Tomasz


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/