x86_64: prepare shared kernel/pci-dma.c
authorThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:14:11 +0000 (11:14 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Thu, 11 Oct 2007 09:14:11 +0000 (11:14 +0200)
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86_64/kernel/Makefile
arch/x86_64/kernel/pci-dma.c [deleted file]
arch/x86_64/kernel/pci-dma_64.c [new file with mode: 0644]

index d71bbfb..8501605 100644 (file)
@@ -8,7 +8,7 @@ obj-y   := process.o signal.o entry.o traps_64.o irq.o \
                ptrace.o time.o ioport.o ldt.o setup.o i8259.o sys_x86_64.o \
                x8664_ksyms.o i387.o syscall.o vsyscall.o \
                setup64.o bootflag.o e820.o reboot.o quirks.o i8237.o \
-               pci-dma.o pci-nommu.o alternative.o hpet.o tsc.o bugs.o \
+               pci-dma_64.o pci-nommu.o alternative.o hpet.o tsc.o bugs.o \
                perfctr-watchdog.o
 
 obj-$(CONFIG_STACKTRACE)       += stacktrace.o
diff --git a/arch/x86_64/kernel/pci-dma.c b/arch/x86_64/kernel/pci-dma.c
deleted file mode 100644 (file)
index 2971144..0000000
+++ /dev/null
@@ -1,346 +0,0 @@
-/*
- * Dynamic DMA mapping support.
- */
-
-#include <linux/types.h>
-#include <linux/mm.h>
-#include <linux/string.h>
-#include <linux/pci.h>
-#include <linux/module.h>
-#include <asm/io.h>
-#include <asm/iommu.h>
-#include <asm/calgary.h>
-
-int iommu_merge __read_mostly = 0;
-EXPORT_SYMBOL(iommu_merge);
-
-dma_addr_t bad_dma_address __read_mostly;
-EXPORT_SYMBOL(bad_dma_address);
-
-/* This tells the BIO block layer to assume merging. Default to off
-   because we cannot guarantee merging later. */
-int iommu_bio_merge __read_mostly = 0;
-EXPORT_SYMBOL(iommu_bio_merge);
-
-static int iommu_sac_force __read_mostly = 0;
-
-int no_iommu __read_mostly;
-#ifdef CONFIG_IOMMU_DEBUG
-int panic_on_overflow __read_mostly = 1;
-int force_iommu __read_mostly = 1;
-#else
-int panic_on_overflow __read_mostly = 0;
-int force_iommu __read_mostly= 0;
-#endif
-
-/* Set this to 1 if there is a HW IOMMU in the system */
-int iommu_detected __read_mostly = 0;
-
-/* Dummy device used for NULL arguments (normally ISA). Better would
-   be probably a smaller DMA mask, but this is bug-to-bug compatible
-   to i386. */
-struct device fallback_dev = {
-       .bus_id = "fallback device",
-       .coherent_dma_mask = DMA_32BIT_MASK,
-       .dma_mask = &fallback_dev.coherent_dma_mask,
-};
-
-/* Allocate DMA memory on node near device */
-noinline static void *
-dma_alloc_pages(struct device *dev, gfp_t gfp, unsigned order)
-{
-       struct page *page;
-       int node;
-#ifdef CONFIG_PCI
-       if (dev->bus == &pci_bus_type)
-               node = pcibus_to_node(to_pci_dev(dev)->bus);
-       else
-#endif
-               node = numa_node_id();
-
-       if (node < first_node(node_online_map))
-               node = first_node(node_online_map);
-
-       page = alloc_pages_node(node, gfp, order);
-       return page ? page_address(page) : NULL;
-}
-
-/*
- * Allocate memory for a coherent mapping.
- */
-void *
-dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
-                  gfp_t gfp)
-{
-       void *memory;
-       unsigned long dma_mask = 0;
-       u64 bus;
-
-       if (!dev)
-               dev = &fallback_dev;
-       dma_mask = dev->coherent_dma_mask;
-       if (dma_mask == 0)
-               dma_mask = DMA_32BIT_MASK;
-
-       /* Device not DMA able */
-       if (dev->dma_mask == NULL)
-               return NULL;
-
-       /* Don't invoke OOM killer */
-       gfp |= __GFP_NORETRY;
-
-       /* Kludge to make it bug-to-bug compatible with i386. i386
-          uses the normal dma_mask for alloc_coherent. */
-       dma_mask &= *dev->dma_mask;
-
-       /* Why <=? Even when the mask is smaller than 4GB it is often
-          larger than 16MB and in this case we have a chance of
-          finding fitting memory in the next higher zone first. If
-          not retry with true GFP_DMA. -AK */
-       if (dma_mask <= DMA_32BIT_MASK)
-               gfp |= GFP_DMA32;
-
- again:
-       memory = dma_alloc_pages(dev, gfp, get_order(size));
-       if (memory == NULL)
-               return NULL;
-
-       {
-               int high, mmu;
-               bus = virt_to_bus(memory);
-               high = (bus + size) >= dma_mask;
-               mmu = high;
-               if (force_iommu && !(gfp & GFP_DMA))
-                       mmu = 1;
-               else if (high) {
-                       free_pages((unsigned long)memory,
-                                  get_order(size));
-
-                       /* Don't use the 16MB ZONE_DMA unless absolutely
-                          needed. It's better to use remapping first. */
-                       if (dma_mask < DMA_32BIT_MASK && !(gfp & GFP_DMA)) {
-                               gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
-                               goto again;
-                       }
-
-                       /* Let low level make its own zone decisions */
-                       gfp &= ~(GFP_DMA32|GFP_DMA);
-
-                       if (dma_ops->alloc_coherent)
-                               return dma_ops->alloc_coherent(dev, size,
-                                                          dma_handle, gfp);
-                       return NULL;
-               }
-
-               memset(memory, 0, size);
-               if (!mmu) {
-                       *dma_handle = virt_to_bus(memory);
-                       return memory;
-               }
-       }
-
-       if (dma_ops->alloc_coherent) {
-               free_pages((unsigned long)memory, get_order(size));
-               gfp &= ~(GFP_DMA|GFP_DMA32);
-               return dma_ops->alloc_coherent(dev, size, dma_handle, gfp);
-       }
-
-       if (dma_ops->map_simple) {
-               *dma_handle = dma_ops->map_simple(dev, memory,
-                                             size,
-                                             PCI_DMA_BIDIRECTIONAL);
-               if (*dma_handle != bad_dma_address)
-                       return memory;
-       }
-
-       if (panic_on_overflow)
-               panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n",size);
-       free_pages((unsigned long)memory, get_order(size));
-       return NULL;
-}
-EXPORT_SYMBOL(dma_alloc_coherent);
-
-/*
- * Unmap coherent memory.
- * The caller must ensure that the device has finished accessing the mapping.
- */
-void dma_free_coherent(struct device *dev, size_t size,
-                        void *vaddr, dma_addr_t bus)
-{
-       if (dma_ops->unmap_single)
-               dma_ops->unmap_single(dev, bus, size, 0);
-       free_pages((unsigned long)vaddr, get_order(size));
-}
-EXPORT_SYMBOL(dma_free_coherent);
-
-static int forbid_dac __read_mostly;
-
-int dma_supported(struct device *dev, u64 mask)
-{
-#ifdef CONFIG_PCI
-       if (mask > 0xffffffff && forbid_dac > 0) {
-
-
-
-               printk(KERN_INFO "PCI: Disallowing DAC for device %s\n", dev->bus_id);
-               return 0;
-       }
-#endif
-
-       if (dma_ops->dma_supported)
-               return dma_ops->dma_supported(dev, mask);
-
-       /* Copied from i386. Doesn't make much sense, because it will
-          only work for pci_alloc_coherent.
-          The caller just has to use GFP_DMA in this case. */
-        if (mask < DMA_24BIT_MASK)
-                return 0;
-
-       /* Tell the device to use SAC when IOMMU force is on.  This
-          allows the driver to use cheaper accesses in some cases.
-
-          Problem with this is that if we overflow the IOMMU area and
-          return DAC as fallback address the device may not handle it
-          correctly.
-
-          As a special case some controllers have a 39bit address
-          mode that is as efficient as 32bit (aic79xx). Don't force
-          SAC for these.  Assume all masks <= 40 bits are of this
-          type. Normally this doesn't make any difference, but gives
-          more gentle handling of IOMMU overflow. */
-       if (iommu_sac_force && (mask >= DMA_40BIT_MASK)) {
-               printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
-               return 0;
-       }
-
-       return 1;
-}
-EXPORT_SYMBOL(dma_supported);
-
-int dma_set_mask(struct device *dev, u64 mask)
-{
-       if (!dev->dma_mask || !dma_supported(dev, mask))
-               return -EIO;
-       *dev->dma_mask = mask;
-       return 0;
-}
-EXPORT_SYMBOL(dma_set_mask);
-
-/*
- * See <Documentation/x86_64/boot-options.txt> for the iommu kernel parameter
- * documentation.
- */
-__init int iommu_setup(char *p)
-{
-       iommu_merge = 1;
-
-       if (!p)
-               return -EINVAL;
-
-       while (*p) {
-               if (!strncmp(p,"off",3))
-                       no_iommu = 1;
-               /* gart_parse_options has more force support */
-               if (!strncmp(p,"force",5))
-                       force_iommu = 1;
-               if (!strncmp(p,"noforce",7)) {
-                       iommu_merge = 0;
-                       force_iommu = 0;
-               }
-
-               if (!strncmp(p, "biomerge",8)) {
-                       iommu_bio_merge = 4096;
-                       iommu_merge = 1;
-                       force_iommu = 1;
-               }
-               if (!strncmp(p, "panic",5))
-                       panic_on_overflow = 1;
-               if (!strncmp(p, "nopanic",7))
-                       panic_on_overflow = 0;
-               if (!strncmp(p, "merge",5)) {
-                       iommu_merge = 1;
-                       force_iommu = 1;
-               }
-               if (!strncmp(p, "nomerge",7))
-                       iommu_merge = 0;
-               if (!strncmp(p, "forcesac",8))
-                       iommu_sac_force = 1;
-               if (!strncmp(p, "allowdac", 8))
-                       forbid_dac = 0;
-               if (!strncmp(p, "nodac", 5))
-                       forbid_dac = -1;
-
-#ifdef CONFIG_SWIOTLB
-               if (!strncmp(p, "soft",4))
-                       swiotlb = 1;
-#endif
-
-#ifdef CONFIG_IOMMU
-               gart_parse_options(p);
-#endif
-
-#ifdef CONFIG_CALGARY_IOMMU
-               if (!strncmp(p, "calgary", 7))
-                       use_calgary = 1;
-#endif /* CONFIG_CALGARY_IOMMU */
-
-               p += strcspn(p, ",");
-               if (*p == ',')
-                       ++p;
-       }
-       return 0;
-}
-early_param("iommu", iommu_setup);
-
-void __init pci_iommu_alloc(void)
-{
-       /*
-        * The order of these functions is important for
-        * fall-back/fail-over reasons
-        */
-#ifdef CONFIG_IOMMU
-       iommu_hole_init();
-#endif
-
-#ifdef CONFIG_CALGARY_IOMMU
-       detect_calgary();
-#endif
-
-#ifdef CONFIG_SWIOTLB
-       pci_swiotlb_init();
-#endif
-}
-
-static int __init pci_iommu_init(void)
-{
-#ifdef CONFIG_CALGARY_IOMMU
-       calgary_iommu_init();
-#endif
-
-#ifdef CONFIG_IOMMU
-       gart_iommu_init();
-#endif
-
-       no_iommu_init();
-       return 0;
-}
-
-void pci_iommu_shutdown(void)
-{
-       gart_iommu_shutdown();
-}
-
-#ifdef CONFIG_PCI
-/* Many VIA bridges seem to corrupt data for DAC. Disable it here */
-
-static __devinit void via_no_dac(struct pci_dev *dev)
-{
-       if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI && forbid_dac == 0) {
-               printk(KERN_INFO "PCI: VIA PCI bridge detected. Disabling DAC.\n");
-               forbid_dac = 1;
-       }
-}
-DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_VIA, PCI_ANY_ID, via_no_dac);
-#endif
-/* Must execute after PCI subsystem */
-fs_initcall(pci_iommu_init);
diff --git a/arch/x86_64/kernel/pci-dma_64.c b/arch/x86_64/kernel/pci-dma_64.c
new file mode 100644 (file)
index 0000000..2971144
--- /dev/null
@@ -0,0 +1,346 @@
+/*
+ * Dynamic DMA mapping support.
+ */
+
+#include <linux/types.h>
+#include <linux/mm.h>
+#include <linux/string.h>
+#include <linux/pci.h>
+#include <linux/module.h>
+#include <asm/io.h>
+#include <asm/iommu.h>
+#include <asm/calgary.h>
+
+int iommu_merge __read_mostly = 0;
+EXPORT_SYMBOL(iommu_merge);
+
+dma_addr_t bad_dma_address __read_mostly;
+EXPORT_SYMBOL(bad_dma_address);
+
+/* This tells the BIO block layer to assume merging. Default to off
+   because we cannot guarantee merging later. */
+int iommu_bio_merge __read_mostly = 0;
+EXPORT_SYMBOL(iommu_bio_merge);
+
+static int iommu_sac_force __read_mostly = 0;
+
+int no_iommu __read_mostly;
+#ifdef CONFIG_IOMMU_DEBUG
+int panic_on_overflow __read_mostly = 1;
+int force_iommu __read_mostly = 1;
+#else
+int panic_on_overflow __read_mostly = 0;
+int force_iommu __read_mostly= 0;
+#endif
+
+/* Set this to 1 if there is a HW IOMMU in the system */
+int iommu_detected __read_mostly = 0;
+
+/* Dummy device used for NULL arguments (normally ISA). Better would
+   be probably a smaller DMA mask, but this is bug-to-bug compatible
+   to i386. */
+struct device fallback_dev = {
+       .bus_id = "fallback device",
+       .coherent_dma_mask = DMA_32BIT_MASK,
+       .dma_mask = &fallback_dev.coherent_dma_mask,
+};
+
+/* Allocate DMA memory on node near device */
+noinline static void *
+dma_alloc_pages(struct device *dev, gfp_t gfp, unsigned order)
+{
+       struct page *page;
+       int node;
+#ifdef CONFIG_PCI
+       if (dev->bus == &pci_bus_type)
+               node = pcibus_to_node(to_pci_dev(dev)->bus);
+       else
+#endif
+               node = numa_node_id();
+
+       if (node < first_node(node_online_map))
+               node = first_node(node_online_map);
+
+       page = alloc_pages_node(node, gfp, order);
+       return page ? page_address(page) : NULL;
+}
+
+/*
+ * Allocate memory for a coherent mapping.
+ */
+void *
+dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *dma_handle,
+                  gfp_t gfp)
+{
+       void *memory;
+       unsigned long dma_mask = 0;
+       u64 bus;
+
+       if (!dev)
+               dev = &fallback_dev;
+       dma_mask = dev->coherent_dma_mask;
+       if (dma_mask == 0)
+               dma_mask = DMA_32BIT_MASK;
+
+       /* Device not DMA able */
+       if (dev->dma_mask == NULL)
+               return NULL;
+
+       /* Don't invoke OOM killer */
+       gfp |= __GFP_NORETRY;
+
+       /* Kludge to make it bug-to-bug compatible with i386. i386
+          uses the normal dma_mask for alloc_coherent. */
+       dma_mask &= *dev->dma_mask;
+
+       /* Why <=? Even when the mask is smaller than 4GB it is often
+          larger than 16MB and in this case we have a chance of
+          finding fitting memory in the next higher zone first. If
+          not retry with true GFP_DMA. -AK */
+       if (dma_mask <= DMA_32BIT_MASK)
+               gfp |= GFP_DMA32;
+
+ again:
+       memory = dma_alloc_pages(dev, gfp, get_order(size));
+       if (memory == NULL)
+               return NULL;
+
+       {
+               int high, mmu;
+               bus = virt_to_bus(memory);
+               high = (bus + size) >= dma_mask;
+               mmu = high;
+               if (force_iommu && !(gfp & GFP_DMA))
+                       mmu = 1;
+               else if (high) {
+                       free_pages((unsigned long)memory,
+                                  get_order(size));
+
+                       /* Don't use the 16MB ZONE_DMA unless absolutely
+                          needed. It's better to use remapping first. */
+                       if (dma_mask < DMA_32BIT_MASK && !(gfp & GFP_DMA)) {
+                               gfp = (gfp & ~GFP_DMA32) | GFP_DMA;
+                               goto again;
+                       }
+
+                       /* Let low level make its own zone decisions */
+                       gfp &= ~(GFP_DMA32|GFP_DMA);
+
+                       if (dma_ops->alloc_coherent)
+                               return dma_ops->alloc_coherent(dev, size,
+                                                          dma_handle, gfp);
+                       return NULL;
+               }
+
+               memset(memory, 0, size);
+               if (!mmu) {
+                       *dma_handle = virt_to_bus(memory);
+                       return memory;
+               }
+       }
+
+       if (dma_ops->alloc_coherent) {
+               free_pages((unsigned long)memory, get_order(size));
+               gfp &= ~(GFP_DMA|GFP_DMA32);
+               return dma_ops->alloc_coherent(dev, size, dma_handle, gfp);
+       }
+
+       if (dma_ops->map_simple) {
+               *dma_handle = dma_ops->map_simple(dev, memory,
+                                             size,
+                                             PCI_DMA_BIDIRECTIONAL);
+               if (*dma_handle != bad_dma_address)
+                       return memory;
+       }
+
+       if (panic_on_overflow)
+               panic("dma_alloc_coherent: IOMMU overflow by %lu bytes\n",size);
+       free_pages((unsigned long)memory, get_order(size));
+       return NULL;
+}
+EXPORT_SYMBOL(dma_alloc_coherent);
+
+/*
+ * Unmap coherent memory.
+ * The caller must ensure that the device has finished accessing the mapping.
+ */
+void dma_free_coherent(struct device *dev, size_t size,
+                        void *vaddr, dma_addr_t bus)
+{
+       if (dma_ops->unmap_single)
+               dma_ops->unmap_single(dev, bus, size, 0);
+       free_pages((unsigned long)vaddr, get_order(size));
+}
+EXPORT_SYMBOL(dma_free_coherent);
+
+static int forbid_dac __read_mostly;
+
+int dma_supported(struct device *dev, u64 mask)
+{
+#ifdef CONFIG_PCI
+       if (mask > 0xffffffff && forbid_dac > 0) {
+
+
+
+               printk(KERN_INFO "PCI: Disallowing DAC for device %s\n", dev->bus_id);
+               return 0;
+       }
+#endif
+
+       if (dma_ops->dma_supported)
+               return dma_ops->dma_supported(dev, mask);
+
+       /* Copied from i386. Doesn't make much sense, because it will
+          only work for pci_alloc_coherent.
+          The caller just has to use GFP_DMA in this case. */
+        if (mask < DMA_24BIT_MASK)
+                return 0;
+
+       /* Tell the device to use SAC when IOMMU force is on.  This
+          allows the driver to use cheaper accesses in some cases.
+
+          Problem with this is that if we overflow the IOMMU area and
+          return DAC as fallback address the device may not handle it
+          correctly.
+
+          As a special case some controllers have a 39bit address
+          mode that is as efficient as 32bit (aic79xx). Don't force
+          SAC for these.  Assume all masks <= 40 bits are of this
+          type. Normally this doesn't make any difference, but gives
+          more gentle handling of IOMMU overflow. */
+       if (iommu_sac_force && (mask >= DMA_40BIT_MASK)) {
+               printk(KERN_INFO "%s: Force SAC with mask %Lx\n", dev->bus_id,mask);
+               return 0;
+       }
+
+       return 1;
+}
+EXPORT_SYMBOL(dma_supported);
+
+int dma_set_mask(struct device *dev, u64 mask)
+{
+       if (!dev->dma_mask || !dma_supported(dev, mask))
+               return -EIO;
+       *dev->dma_mask = mask;
+       return 0;
+}
+EXPORT_SYMBOL(dma_set_mask);
+
+/*
+ * See <Documentation/x86_64/boot-options.txt> for the iommu kernel parameter
+ * documentation.
+ */
+__init int iommu_setup(char *p)
+{
+       iommu_merge = 1;
+
+       if (!p)
+               return -EINVAL;
+
+       while (*p) {
+               if (!strncmp(p,"off",3))
+                       no_iommu = 1;
+               /* gart_parse_options has more force support */
+               if (!strncmp(p,"force",5))
+                       force_iommu = 1;
+               if (!strncmp(p,"noforce",7)) {
+                       iommu_merge = 0;
+                       force_iommu = 0;
+               }
+
+               if (!strncmp(p, "biomerge",8)) {
+                       iommu_bio_merge = 4096;
+                       iommu_merge = 1;
+                       force_iommu = 1;
+               }
+               if (!strncmp(p, "panic",5))
+                       panic_on_overflow = 1;
+               if (!strncmp(p, "nopanic",7))
+                       panic_on_overflow = 0;
+               if (!strncmp(p, "merge",5)) {
+                       iommu_merge = 1;
+                       force_iommu = 1;
+               }
+               if (!strncmp(p, "nomerge",7))
+                       iommu_merge = 0;
+               if (!strncmp(p, "forcesac",8))
+                       iommu_sac_force = 1;
+               if (!strncmp(p, "allowdac", 8))
+                       forbid_dac = 0;
+               if (!strncmp(p, "nodac", 5))
+                       forbid_dac = -1;
+
+#ifdef CONFIG_SWIOTLB
+               if (!strncmp(p, "soft",4))
+                       swiotlb = 1;
+#endif
+
+#ifdef CONFIG_IOMMU
+               gart_parse_options(p);
+#endif
+
+#ifdef CONFIG_CALGARY_IOMMU
+               if (!strncmp(p, "calgary", 7))
+                       use_calgary = 1;
+#endif /* CONFIG_CALGARY_IOMMU */
+
+               p += strcspn(p, ",");
+               if (*p == ',')
+                       ++p;
+       }
+       return 0;
+}
+early_param("iommu", iommu_setup);
+
+void __init pci_iommu_alloc(void)
+{
+       /*
+        * The order of these functions is important for
+        * fall-back/fail-over reasons
+        */
+#ifdef CONFIG_IOMMU
+       iommu_hole_init();
+#endif
+
+#ifdef CONFIG_CALGARY_IOMMU
+       detect_calgary();
+#endif
+
+#ifdef CONFIG_SWIOTLB
+       pci_swiotlb_init();
+#endif
+}
+
+static int __init pci_iommu_init(void)
+{
+#ifdef CONFIG_CALGARY_IOMMU
+       calgary_iommu_init();
+#endif
+
+#ifdef CONFIG_IOMMU
+       gart_iommu_init();
+#endif
+
+       no_iommu_init();
+       return 0;
+}
+
+void pci_iommu_shutdown(void)
+{
+       gart_iommu_shutdown();
+}
+
+#ifdef CONFIG_PCI
+/* Many VIA bridges seem to corrupt data for DAC. Disable it here */
+
+static __devinit void via_no_dac(struct pci_dev *dev)
+{
+       if ((dev->class >> 8) == PCI_CLASS_BRIDGE_PCI && forbid_dac == 0) {
+               printk(KERN_INFO "PCI: VIA PCI bridge detected. Disabling DAC.\n");
+               forbid_dac = 1;
+       }
+}
+DECLARE_PCI_FIXUP_FINAL(PCI_VENDOR_ID_VIA, PCI_ANY_ID, via_no_dac);
+#endif
+/* Must execute after PCI subsystem */
+fs_initcall(pci_iommu_init);