powerpc/fsl-pci: Allow 64-bit PCIe devices to DMA to any memory address

There is an issue on FSL-BookE 64-bit devices (P5020) in which PCIe
devices that are capable of doing 64-bit DMAs (like an Intel e1000) do
not function and crash the kernel if we have >4G of memory in the system.

The reason is that the existing code only sets up one inbound window for
access to system memory across PCIe.  That window is limited to a 32-bit
address space.  So on systems we'll end up utilizing SWIOTLB for dma
mappings.  However SWIOTLB dma ops implement dma_alloc_coherent() as
dma_direct_alloc_coherent().  Thus we can end up with dma addresses that
are not accessible because of the inbound window limitation.

We could possibly set the SWIOTLB alloc_coherent op to
swiotlb_alloc_coherent() however that does not address the issue since
the swiotlb_alloc_coherent() will behave almost identical to
dma_direct_alloc_coherent() since the devices coherent_dma_mask will be
greater than any address allocated by swiotlb_alloc_coherent() and thus
we'll never bounce buffer it into a range that would be dma-able.

The easiest and best solution is to just make it so that a 64-bit
capable device is able to DMA to any internal system address.

We accomplish this by opening up a second inbound window that maps all
of memory above the internal SoC address width so we can set it up to
access all of the internal SoC address space if needed.

We than fixup the dma_ops and dma_offset for PCIe devices with a dma
mask greater than the maximum internal SoC address.

Signed-off-by: Kumar Gala <galak@kernel.crashing.org>

diff --git a/arch/powerpc/sysdev/fsl_pci.c b/arch/powerpc/sysdev/fsl_pci.c
index 4ce547e00473ef6a18206048425dd5aee01e3ec5..8f92446ff0ae1f065d0feca9267a8125ed7d5a41 100644 (file)
--- a/arch/powerpc/sysdev/fsl_pci.c
+++ b/arch/powerpc/sysdev/fsl_pci.c
@@ -65,6 +65,30 @@ static int __init fsl_pcie_check_link(struct pci_controller *hose)
 }
 
 #if defined(CONFIG_FSL_SOC_BOOKE) || defined(CONFIG_PPC_86xx)
+
+#define MAX_PHYS_ADDR_BITS     40
+static u64 pci64_dma_offset = 1ull << MAX_PHYS_ADDR_BITS;
+
+static int fsl_pci_dma_set_mask(struct device *dev, u64 dma_mask)
+{
+       if (!dev->dma_mask || !dma_supported(dev, dma_mask))
+               return -EIO;
+
+       /*
+        * Fixup PCI devices that are able to DMA to above the physical
+        * address width of the SoC such that we can address any internal
+        * SoC address from across PCI if needed
+        */
+       if ((dev->bus == &pci_bus_type) &&
+           dma_mask >= DMA_BIT_MASK(MAX_PHYS_ADDR_BITS)) {
+               set_dma_ops(dev, &dma_direct_ops);
+               set_dma_offset(dev, pci64_dma_offset);
+       }
+
+       *dev->dma_mask = dma_mask;
+       return 0;
+}
+
 static int __init setup_one_atmu(struct ccsr_pci __iomem *pci,
        unsigned int index, const struct resource *res,
        resource_size_t offset)
@@ -228,6 +252,37 @@ static void __init setup_pci_atmu(struct pci_controller *hose,
 
                hose->dma_window_base_cur = 0x00000000;
                hose->dma_window_size = (resource_size_t)sz;
+
+               /*
+                * if we have >4G of memory setup second PCI inbound window to
+                * let devices that are 64-bit address capable to work w/o
+                * SWIOTLB and access the full range of memory
+                */
+               if (sz != mem) {
+                       mem_log = __ilog2_u64(mem);
+
+                       /* Size window up if we dont fit in exact power-of-2 */
+                       if ((1ull << mem_log) != mem)
+                               mem_log++;
+
+                       piwar = (piwar & ~PIWAR_SZ_MASK) | (mem_log - 1);
+
+                       /* Setup inbound memory window */
+                       out_be32(&pci->piw[win_idx].pitar,  0x00000000);
+                       out_be32(&pci->piw[win_idx].piwbear,
+                                       pci64_dma_offset >> 44);
+                       out_be32(&pci->piw[win_idx].piwbar,
+                                       pci64_dma_offset >> 12);
+                       out_be32(&pci->piw[win_idx].piwar,  piwar);
+
+                       /*
+                        * install our own dma_set_mask handler to fixup dma_ops
+                        * and dma_offset
+                        */
+                       ppc_md.dma_set_mask = fsl_pci_dma_set_mask;
+
+                       pr_info("%s: Setup 64-bit PCI DMA window\n", name);
+               }
        } else {
                u64 paddr = 0;