[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20171219101918.GA18921@red-moon>
Date: Tue, 19 Dec 2017 10:19:18 +0000
From: Lorenzo Pieralisi <lorenzo.pieralisi@....com>
To: Niklas Cassel <niklas.cassel@...s.com>
Cc: Jingoo Han <jingoohan1@...il.com>,
Joao Pinto <Joao.Pinto@...opsys.com>,
Bjorn Helgaas <bhelgaas@...gle.com>,
Niklas Cassel <niklass@...s.com>, linux-pci@...r.kernel.org,
linux-kernel@...r.kernel.org
Subject: Re: [PATCH v5 01/18] PCI: dwc: Use the DMA-API to get the MSI address
On Mon, Nov 20, 2017 at 02:32:04PM +0100, Niklas Cassel wrote:
> Use the DMA-API to get the MSI address. This address will be written to
> our PCI config space and to the register which determines which AXI
> address the DWC IP will spoof for incoming MSI irqs.
>
> Since it is a PCIe endpoint device, rather than the CPU, that is supposed
> to write to the MSI address, the proper way to get the MSI address is by
> using the DMA API, not by using virt_to_phys().
>
> Using virt_to_phys() might work on some systems, but using the DMA API
> should work on all systems.
>
> This is essentially the same thing as allocating a buffer in a driver
> to which the endpoint will write to. To do this, we use the DMA API.
>
> Signed-off-by: Niklas Cassel <niklas.cassel@...s.com>
> ---
> drivers/pci/dwc/pcie-designware-host.c | 15 ++++++++++++---
> drivers/pci/dwc/pcie-designware.h | 3 ++-
> 2 files changed, 14 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/pci/dwc/pcie-designware-host.c b/drivers/pci/dwc/pcie-designware-host.c
> index 81e2157a7cfb..33b52fe98a01 100644
> --- a/drivers/pci/dwc/pcie-designware-host.c
> +++ b/drivers/pci/dwc/pcie-designware-host.c
> @@ -83,10 +83,19 @@ irqreturn_t dw_handle_msi_irq(struct pcie_port *pp)
>
> void dw_pcie_msi_init(struct pcie_port *pp)
> {
> + struct dw_pcie *pci = to_dw_pcie_from_pp(pp);
> + struct device *dev = pci->dev;
> + struct page *page;
> u64 msi_target;
>
> - pp->msi_data = __get_free_pages(GFP_KERNEL, 0);
> - msi_target = virt_to_phys((void *)pp->msi_data);
> + page = alloc_page(GFP_KERNEL | GFP_DMA32);
See this thread about GFP_DMA32:
https://patchwork.ozlabs.org/patch/834864/
I need to look back at this set earlier versions, I do not know why
you change the allocation flags but GFP_DMA32 may not provide what
you need - I think you should either remove it or provide a
justification for it given the discussion above.
Lorenzo
> + pp->msi_data = dma_map_page(dev, page, 0, PAGE_SIZE, DMA_FROM_DEVICE);
> + if (dma_mapping_error(dev, pp->msi_data)) {
> + dev_err(dev, "failed to map MSI data\n");
> + __free_page(page);
> + return;
> + }
> + msi_target = (u64)pp->msi_data;
>
> /* program the msi_data */
> dw_pcie_wr_own_conf(pp, PCIE_MSI_ADDR_LO, 4,
> @@ -187,7 +196,7 @@ static void dw_msi_setup_msg(struct pcie_port *pp, unsigned int irq, u32 pos)
> if (pp->ops->get_msi_addr)
> msi_target = pp->ops->get_msi_addr(pp);
> else
> - msi_target = virt_to_phys((void *)pp->msi_data);
> + msi_target = (u64)pp->msi_data;
>
> msg.address_lo = (u32)(msi_target & 0xffffffff);
> msg.address_hi = (u32)(msi_target >> 32 & 0xffffffff);
> diff --git a/drivers/pci/dwc/pcie-designware.h b/drivers/pci/dwc/pcie-designware.h
> index e5d9d77b778e..ecdede68522a 100644
> --- a/drivers/pci/dwc/pcie-designware.h
> +++ b/drivers/pci/dwc/pcie-designware.h
> @@ -14,6 +14,7 @@
> #ifndef _PCIE_DESIGNWARE_H
> #define _PCIE_DESIGNWARE_H
>
> +#include <linux/dma-mapping.h>
> #include <linux/irq.h>
> #include <linux/msi.h>
> #include <linux/pci.h>
> @@ -168,7 +169,7 @@ struct pcie_port {
> const struct dw_pcie_host_ops *ops;
> int msi_irq;
> struct irq_domain *irq_domain;
> - unsigned long msi_data;
> + dma_addr_t msi_data;
> DECLARE_BITMAP(msi_irq_in_use, MAX_MSI_IRQS);
> };
>
> --
> 2.14.2
>
Powered by blists - more mailing lists