Intel IOMMU: Intel IOMMU driver
Keshavamurthy, Anil S [Sun, 21 Oct 2007 23:41:49 +0000 (16:41 -0700)]
Actual intel IOMMU driver.  Hardware spec can be found at:
http://www.intel.com/technology/virtualization

This driver sets X86_64 'dma_ops', so hook into standard DMA APIs.  In this
way, PCI driver will get virtual DMA address.  This change is transparent to
PCI drivers.

[akpm@linux-foundation.org: remove unneeded cast]
[akpm@linux-foundation.org: build fix]
[bunk@stusta.de: fix duplicate CONFIG_DMAR Makefile line]
Signed-off-by: Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
Cc: Andi Kleen <ak@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Muli Ben-Yehuda <muli@il.ibm.com>
Cc: "Siddha, Suresh B" <suresh.b.siddha@intel.com>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Ashok Raj <ashok.raj@intel.com>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Christoph Lameter <clameter@sgi.com>
Cc: Greg KH <greg@kroah.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Documentation/Intel-IOMMU.txt [new file with mode: 0644]
Documentation/kernel-parameters.txt
arch/x86/kernel/pci-dma_64.c
drivers/pci/Makefile
drivers/pci/intel-iommu.c [new file with mode: 0644]
drivers/pci/intel-iommu.h [new file with mode: 0644]
include/linux/dmar.h

diff --git a/Documentation/Intel-IOMMU.txt b/Documentation/Intel-IOMMU.txt
new file mode 100644 (file)
index 0000000..cbb4dba
--- /dev/null
@@ -0,0 +1,93 @@
+Linux IOMMU Support
+===================
+
+The architecture spec can be obtained from the below location.
+
+http://www.intel.com/technology/virtualization/
+
+This guide gives a quick cheat sheet for some basic understanding.
+
+Some Keywords
+
+DMAR - DMA remapping
+DRHD - DMA Engine Reporting Structure
+RMRR - Reserved memory Region Reporting Structure
+ZLR  - Zero length reads from PCI devices
+IOVA - IO Virtual address.
+
+Basic stuff
+-----------
+
+ACPI enumerates and lists the different DMA engines in the platform, and
+device scope relationships between PCI devices and which DMA engine  controls
+them.
+
+What is RMRR?
+-------------
+
+There are some devices the BIOS controls, for e.g USB devices to perform
+PS2 emulation. The regions of memory used for these devices are marked
+reserved in the e820 map. When we turn on DMA translation, DMA to those
+regions will fail. Hence BIOS uses RMRR to specify these regions along with
+devices that need to access these regions. OS is expected to setup
+unity mappings for these regions for these devices to access these regions.
+
+How is IOVA generated?
+---------------------
+
+Well behaved drivers call pci_map_*() calls before sending command to device
+that needs to perform DMA. Once DMA is completed and mapping is no longer
+required, device performs a pci_unmap_*() calls to unmap the region.
+
+The Intel IOMMU driver allocates a virtual address per domain. Each PCIE
+device has its own domain (hence protection). Devices under p2p bridges
+share the virtual address with all devices under the p2p bridge due to
+transaction id aliasing for p2p bridges.
+
+IOVA generation is pretty generic. We used the same technique as vmalloc()
+but these are not global address spaces, but separate for each domain.
+Different DMA engines may support different number of domains.
+
+We also allocate gaurd pages with each mapping, so we can attempt to catch
+any overflow that might happen.
+
+
+Graphics Problems?
+------------------
+If you encounter issues with graphics devices, you can try adding
+option intel_iommu=igfx_off to turn off the integrated graphics engine.
+
+Some exceptions to IOVA
+-----------------------
+Interrupt ranges are not address translated, (0xfee00000 - 0xfeefffff).
+The same is true for peer to peer transactions. Hence we reserve the
+address from PCI MMIO ranges so they are not allocated for IOVA addresses.
+
+Boot Message Sample
+-------------------
+
+Something like this gets printed indicating presence of DMAR tables
+in ACPI.
+
+ACPI: DMAR (v001 A M I  OEMDMAR  0x00000001 MSFT 0x00000097) @ 0x000000007f5b5ef0
+
+When DMAR is being processed and initialized by ACPI, prints DMAR locations
+and any RMRR's processed.
+
+ACPI DMAR:Host address width 36
+ACPI DMAR:DRHD (flags: 0x00000000)base: 0x00000000fed90000
+ACPI DMAR:DRHD (flags: 0x00000000)base: 0x00000000fed91000
+ACPI DMAR:DRHD (flags: 0x00000001)base: 0x00000000fed93000
+ACPI DMAR:RMRR base: 0x00000000000ed000 end: 0x00000000000effff
+ACPI DMAR:RMRR base: 0x000000007f600000 end: 0x000000007fffffff
+
+When DMAR is enabled for use, you will notice..
+
+PCI-DMA: Using DMAR IOMMU
+
+TBD
+----
+
+- For compatibility testing, could use unity map domain for all devices, just
+  provide a 1-1 for all useful memory under a single domain for all devices.
+- API for paravirt ops for abstracting functionlity for VMM folks.
index 6accd36..8157417 100644 (file)
@@ -772,6 +772,16 @@ and is between 256 and 4096 characters. It is defined in the file
 
        inttest=        [IA64]
 
+       intel_iommu=    [DMAR] Intel IOMMU driver (DMAR) option
+               off
+                       Disable intel iommu driver.
+               igfx_off [Default Off]
+                       By default, gfx is mapped as normal device. If a gfx
+                       device has a dedicated DMAR unit, the DMAR unit is
+                       bypassed by not enabling DMAR with this option. In
+                       this case, gfx device will use physical address for
+                       DMA.
+
        io7=            [HW] IO7 for Marvel based alpha systems
                        See comment before marvel_specify_io7 in
                        arch/alpha/kernel/core_marvel.c.
index afaf9f1..393e272 100644 (file)
@@ -7,6 +7,7 @@
 #include <linux/string.h>
 #include <linux/pci.h>
 #include <linux/module.h>
+#include <linux/dmar.h>
 #include <asm/io.h>
 #include <asm/iommu.h>
 #include <asm/calgary.h>
@@ -305,6 +306,8 @@ void __init pci_iommu_alloc(void)
        detect_calgary();
 #endif
 
+       detect_intel_iommu();
+
 #ifdef CONFIG_SWIOTLB
        pci_swiotlb_init();
 #endif
@@ -316,6 +319,8 @@ static int __init pci_iommu_init(void)
        calgary_iommu_init();
 #endif
 
+       intel_iommu_init();
+
 #ifdef CONFIG_IOMMU
        gart_iommu_init();
 #endif
index 836ab2f..5550556 100644 (file)
@@ -21,7 +21,7 @@ obj-$(CONFIG_PCI_MSI) += msi.o
 obj-$(CONFIG_HT_IRQ) += htirq.o
 
 # Build Intel IOMMU support
-obj-$(CONFIG_DMAR) += dmar.o
+obj-$(CONFIG_DMAR) += dmar.o iova.o intel-iommu.o
 
 #
 # Some architectures use the generic PCI setup functions
diff --git a/drivers/pci/intel-iommu.c b/drivers/pci/intel-iommu.c
new file mode 100644 (file)
index 0000000..93ed771
--- /dev/null
@@ -0,0 +1,1957 @@
+/*
+ * Copyright (c) 2006, Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+ * Place - Suite 330, Boston, MA 02111-1307 USA.
+ *
+ * Copyright (C) Ashok Raj <ashok.raj@intel.com>
+ * Copyright (C) Shaohua Li <shaohua.li@intel.com>
+ * Copyright (C) Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
+ */
+
+#include <linux/init.h>
+#include <linux/bitmap.h>
+#include <linux/slab.h>
+#include <linux/irq.h>
+#include <linux/interrupt.h>
+#include <linux/sysdev.h>
+#include <linux/spinlock.h>
+#include <linux/pci.h>
+#include <linux/dmar.h>
+#include <linux/dma-mapping.h>
+#include <linux/mempool.h>
+#include "iova.h"
+#include "intel-iommu.h"
+#include <asm/proto.h> /* force_iommu in this header in x86-64*/
+#include <asm/cacheflush.h>
+#include <asm/iommu.h>
+#include "pci.h"
+
+#define IS_GFX_DEVICE(pdev) ((pdev->class >> 16) == PCI_BASE_CLASS_DISPLAY)
+#define IS_ISA_DEVICE(pdev) ((pdev->class >> 8) == PCI_CLASS_BRIDGE_ISA)
+
+#define IOAPIC_RANGE_START     (0xfee00000)
+#define IOAPIC_RANGE_END       (0xfeefffff)
+#define IOVA_START_ADDR                (0x1000)
+
+#define DEFAULT_DOMAIN_ADDRESS_WIDTH 48
+
+#define DMAR_OPERATION_TIMEOUT (HZ*60) /* 1m */
+
+#define DOMAIN_MAX_ADDR(gaw) ((((u64)1) << gaw) - 1)
+
+static void domain_remove_dev_info(struct dmar_domain *domain);
+
+static int dmar_disabled;
+static int __initdata dmar_map_gfx = 1;
+
+#define DUMMY_DEVICE_DOMAIN_INFO ((struct device_domain_info *)(-1))
+static DEFINE_SPINLOCK(device_domain_lock);
+static LIST_HEAD(device_domain_list);
+
+static int __init intel_iommu_setup(char *str)
+{
+       if (!str)
+               return -EINVAL;
+       while (*str) {
+               if (!strncmp(str, "off", 3)) {
+                       dmar_disabled = 1;
+                       printk(KERN_INFO"Intel-IOMMU: disabled\n");
+               } else if (!strncmp(str, "igfx_off", 8)) {
+                       dmar_map_gfx = 0;
+                       printk(KERN_INFO
+                               "Intel-IOMMU: disable GFX device mapping\n");
+               }
+
+               str += strcspn(str, ",");
+               while (*str == ',')
+                       str++;
+       }
+       return 0;
+}
+__setup("intel_iommu=", intel_iommu_setup);
+
+static struct kmem_cache *iommu_domain_cache;
+static struct kmem_cache *iommu_devinfo_cache;
+static struct kmem_cache *iommu_iova_cache;
+
+static inline void *alloc_pgtable_page(void)
+{
+       return (void *)get_zeroed_page(GFP_ATOMIC);
+}
+
+static inline void free_pgtable_page(void *vaddr)
+{
+       free_page((unsigned long)vaddr);
+}
+
+static inline void *alloc_domain_mem(void)
+{
+       return kmem_cache_alloc(iommu_domain_cache, GFP_ATOMIC);
+}
+
+static inline void free_domain_mem(void *vaddr)
+{
+       kmem_cache_free(iommu_domain_cache, vaddr);
+}
+
+static inline void * alloc_devinfo_mem(void)
+{
+       return kmem_cache_alloc(iommu_devinfo_cache, GFP_ATOMIC);
+}
+
+static inline void free_devinfo_mem(void *vaddr)
+{
+       kmem_cache_free(iommu_devinfo_cache, vaddr);
+}
+
+struct iova *alloc_iova_mem(void)
+{
+       return kmem_cache_alloc(iommu_iova_cache, GFP_ATOMIC);
+}
+
+void free_iova_mem(struct iova *iova)
+{
+       kmem_cache_free(iommu_iova_cache, iova);
+}
+
+static inline void __iommu_flush_cache(
+       struct intel_iommu *iommu, void *addr, int size)
+{
+       if (!ecap_coherent(iommu->ecap))
+               clflush_cache_range(addr, size);
+}
+
+/* Gets context entry for a given bus and devfn */
+static struct context_entry * device_to_context_entry(struct intel_iommu *iommu,
+               u8 bus, u8 devfn)
+{
+       struct root_entry *root;
+       struct context_entry *context;
+       unsigned long phy_addr;
+       unsigned long flags;
+
+       spin_lock_irqsave(&iommu->lock, flags);
+       root = &iommu->root_entry[bus];
+       context = get_context_addr_from_root(root);
+       if (!context) {
+               context = (struct context_entry *)alloc_pgtable_page();
+               if (!context) {
+                       spin_unlock_irqrestore(&iommu->lock, flags);
+                       return NULL;
+               }
+               __iommu_flush_cache(iommu, (void *)context, PAGE_SIZE_4K);
+               phy_addr = virt_to_phys((void *)context);
+               set_root_value(root, phy_addr);
+               set_root_present(root);
+               __iommu_flush_cache(iommu, root, sizeof(*root));
+       }
+       spin_unlock_irqrestore(&iommu->lock, flags);
+       return &context[devfn];
+}
+
+static int device_context_mapped(struct intel_iommu *iommu, u8 bus, u8 devfn)
+{
+       struct root_entry *root;
+       struct context_entry *context;
+       int ret;
+       unsigned long flags;
+
+       spin_lock_irqsave(&iommu->lock, flags);
+       root = &iommu->root_entry[bus];
+       context = get_context_addr_from_root(root);
+       if (!context) {
+               ret = 0;
+               goto out;
+       }
+       ret = context_present(context[devfn]);
+out:
+       spin_unlock_irqrestore(&iommu->lock, flags);
+       return ret;
+}
+
+static void clear_context_table(struct intel_iommu *iommu, u8 bus, u8 devfn)
+{
+       struct root_entry *root;
+       struct context_entry *context;
+       unsigned long flags;
+
+       spin_lock_irqsave(&iommu->lock, flags);
+       root = &iommu->root_entry[bus];
+       context = get_context_addr_from_root(root);
+       if (context) {
+               context_clear_entry(context[devfn]);
+               __iommu_flush_cache(iommu, &context[devfn], \
+                       sizeof(*context));
+       }
+       spin_unlock_irqrestore(&iommu->lock, flags);
+}
+
+static void free_context_table(struct intel_iommu *iommu)
+{
+       struct root_entry *root;
+       int i;
+       unsigned long flags;
+       struct context_entry *context;
+
+       spin_lock_irqsave(&iommu->lock, flags);
+       if (!iommu->root_entry) {
+               goto out;
+       }
+       for (i = 0; i < ROOT_ENTRY_NR; i++) {
+               root = &iommu->root_entry[i];
+               context = get_context_addr_from_root(root);
+               if (context)
+                       free_pgtable_page(context);
+       }
+       free_pgtable_page(iommu->root_entry);
+       iommu->root_entry = NULL;
+out:
+       spin_unlock_irqrestore(&iommu->lock, flags);
+}
+
+/* page table handling */
+#define LEVEL_STRIDE           (9)
+#define LEVEL_MASK             (((u64)1 << LEVEL_STRIDE) - 1)
+
+static inline int agaw_to_level(int agaw)
+{
+       return agaw + 2;
+}
+
+static inline int agaw_to_width(int agaw)
+{
+       return 30 + agaw * LEVEL_STRIDE;
+
+}
+
+static inline int width_to_agaw(int width)
+{
+       return (width - 30) / LEVEL_STRIDE;
+}
+
+static inline unsigned int level_to_offset_bits(int level)
+{
+       return (12 + (level - 1) * LEVEL_STRIDE);
+}
+
+static inline int address_level_offset(u64 addr, int level)
+{
+       return ((addr >> level_to_offset_bits(level)) & LEVEL_MASK);
+}
+
+static inline u64 level_mask(int level)
+{
+       return ((u64)-1 << level_to_offset_bits(level));
+}
+
+static inline u64 level_size(int level)
+{
+       return ((u64)1 << level_to_offset_bits(level));
+}
+
+static inline u64 align_to_level(u64 addr, int level)
+{
+       return ((addr + level_size(level) - 1) & level_mask(level));
+}
+
+static struct dma_pte * addr_to_dma_pte(struct dmar_domain *domain, u64 addr)
+{
+       int addr_width = agaw_to_width(domain->agaw);
+       struct dma_pte *parent, *pte = NULL;
+       int level = agaw_to_level(domain->agaw);
+       int offset;
+       unsigned long flags;
+
+       BUG_ON(!domain->pgd);
+
+       addr &= (((u64)1) << addr_width) - 1;
+       parent = domain->pgd;
+
+       spin_lock_irqsave(&domain->mapping_lock, flags);
+       while (level > 0) {
+               void *tmp_page;
+
+               offset = address_level_offset(addr, level);
+               pte = &parent[offset];
+               if (level == 1)
+                       break;
+
+               if (!dma_pte_present(*pte)) {
+                       tmp_page = alloc_pgtable_page();
+
+                       if (!tmp_page) {
+                               spin_unlock_irqrestore(&domain->mapping_lock,
+                                       flags);
+                               return NULL;
+                       }
+                       __iommu_flush_cache(domain->iommu, tmp_page,
+                                       PAGE_SIZE_4K);
+                       dma_set_pte_addr(*pte, virt_to_phys(tmp_page));
+                       /*
+                        * high level table always sets r/w, last level page
+                        * table control read/write
+                        */
+                       dma_set_pte_readable(*pte);
+                       dma_set_pte_writable(*pte);
+                       __iommu_flush_cache(domain->iommu, pte, sizeof(*pte));
+               }
+               parent = phys_to_virt(dma_pte_addr(*pte));
+               level--;
+       }
+
+       spin_unlock_irqrestore(&domain->mapping_lock, flags);
+       return pte;
+}
+
+/* return address's pte at specific level */
+static struct dma_pte *dma_addr_level_pte(struct dmar_domain *domain, u64 addr,
+               int level)
+{
+       struct dma_pte *parent, *pte = NULL;
+       int total = agaw_to_level(domain->agaw);
+       int offset;
+
+       parent = domain->pgd;
+       while (level <= total) {
+               offset = address_level_offset(addr, total);
+               pte = &parent[offset];
+               if (level == total)
+                       return pte;
+
+               if (!dma_pte_present(*pte))
+                       break;
+               parent = phys_to_virt(dma_pte_addr(*pte));
+               total--;
+       }
+       return NULL;
+}
+
+/* clear one page's page table */
+static void dma_pte_clear_one(struct dmar_domain *domain, u64 addr)
+{
+       struct dma_pte *pte = NULL;
+
+       /* get last level pte */
+       pte = dma_addr_level_pte(domain, addr, 1);
+
+       if (pte) {
+               dma_clear_pte(*pte);
+               __iommu_flush_cache(domain->iommu, pte, sizeof(*pte));
+       }
+}
+
+/* clear last level pte, a tlb flush should be followed */
+static void dma_pte_clear_range(struct dmar_domain *domain, u64 start, u64 end)
+{
+       int addr_width = agaw_to_width(domain->agaw);
+
+       start &= (((u64)1) << addr_width) - 1;
+       end &= (((u64)1) << addr_width) - 1;
+       /* in case it's partial page */
+       start = PAGE_ALIGN_4K(start);
+       end &= PAGE_MASK_4K;
+
+       /* we don't need lock here, nobody else touches the iova range */
+       while (start < end) {
+               dma_pte_clear_one(domain, start);
+               start += PAGE_SIZE_4K;
+       }
+}
+
+/* free page table pages. last level pte should already be cleared */
+static void dma_pte_free_pagetable(struct dmar_domain *domain,
+       u64 start, u64 end)
+{
+       int addr_width = agaw_to_width(domain->agaw);
+       struct dma_pte *pte;
+       int total = agaw_to_level(domain->agaw);
+       int level;
+       u64 tmp;
+
+       start &= (((u64)1) << addr_width) - 1;
+       end &= (((u64)1) << addr_width) - 1;
+
+       /* we don't need lock here, nobody else touches the iova range */
+       level = 2;
+       while (level <= total) {
+               tmp = align_to_level(start, level);
+               if (tmp >= end || (tmp + level_size(level) > end))
+                       return;
+
+               while (tmp < end) {
+                       pte = dma_addr_level_pte(domain, tmp, level);
+                       if (pte) {
+                               free_pgtable_page(
+                                       phys_to_virt(dma_pte_addr(*pte)));
+                               dma_clear_pte(*pte);
+                               __iommu_flush_cache(domain->iommu,
+                                               pte, sizeof(*pte));
+                       }
+                       tmp += level_size(level);
+               }
+               level++;
+       }
+       /* free pgd */
+       if (start == 0 && end >= ((((u64)1) << addr_width) - 1)) {
+               free_pgtable_page(domain->pgd);
+               domain->pgd = NULL;
+       }
+}
+
+/* iommu handling */
+static int iommu_alloc_root_entry(struct intel_iommu *iommu)
+{
+       struct root_entry *root;
+       unsigned long flags;
+
+       root = (struct root_entry *)alloc_pgtable_page();
+       if (!root)
+               return -ENOMEM;
+
+       __iommu_flush_cache(iommu, root, PAGE_SIZE_4K);
+
+       spin_lock_irqsave(&iommu->lock, flags);
+       iommu->root_entry = root;
+       spin_unlock_irqrestore(&iommu->lock, flags);
+
+       return 0;
+}
+
+#define IOMMU_WAIT_OP(iommu, offset, op, cond, sts) \
+{\
+       unsigned long start_time = jiffies;\
+       while (1) {\
+               sts = op (iommu->reg + offset);\
+               if (cond)\
+                       break;\
+               if (time_after(jiffies, start_time + DMAR_OPERATION_TIMEOUT))\
+                       panic("DMAR hardware is malfunctioning\n");\
+               cpu_relax();\
+       }\
+}
+
+static void iommu_set_root_entry(struct intel_iommu *iommu)
+{
+       void *addr;
+       u32 cmd, sts;
+       unsigned long flag;
+
+       addr = iommu->root_entry;
+
+       spin_lock_irqsave(&iommu->register_lock, flag);
+       dmar_writeq(iommu->reg + DMAR_RTADDR_REG, virt_to_phys(addr));
+
+       cmd = iommu->gcmd | DMA_GCMD_SRTP;
+       writel(cmd, iommu->reg + DMAR_GCMD_REG);
+
+       /* Make sure hardware complete it */
+       IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
+               readl, (sts & DMA_GSTS_RTPS), sts);
+
+       spin_unlock_irqrestore(&iommu->register_lock, flag);
+}
+
+static void iommu_flush_write_buffer(struct intel_iommu *iommu)
+{
+       u32 val;
+       unsigned long flag;
+
+       if (!cap_rwbf(iommu->cap))
+               return;
+       val = iommu->gcmd | DMA_GCMD_WBF;
+
+       spin_lock_irqsave(&iommu->register_lock, flag);
+       writel(val, iommu->reg + DMAR_GCMD_REG);
+
+       /* Make sure hardware complete it */
+       IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
+                       readl, (!(val & DMA_GSTS_WBFS)), val);
+
+       spin_unlock_irqrestore(&iommu->register_lock, flag);
+}
+
+/* return value determine if we need a write buffer flush */
+static int __iommu_flush_context(struct intel_iommu *iommu,
+       u16 did, u16 source_id, u8 function_mask, u64 type,
+       int non_present_entry_flush)
+{
+       u64 val = 0;
+       unsigned long flag;
+
+       /*
+        * In the non-present entry flush case, if hardware doesn't cache
+        * non-present entry we do nothing and if hardware cache non-present
+        * entry, we flush entries of domain 0 (the domain id is used to cache
+        * any non-present entries)
+        */
+       if (non_present_entry_flush) {
+               if (!cap_caching_mode(iommu->cap))
+                       return 1;
+               else
+                       did = 0;
+       }
+
+       switch (type) {
+       case DMA_CCMD_GLOBAL_INVL:
+               val = DMA_CCMD_GLOBAL_INVL;
+               break;
+       case DMA_CCMD_DOMAIN_INVL:
+               val = DMA_CCMD_DOMAIN_INVL|DMA_CCMD_DID(did);
+               break;
+       case DMA_CCMD_DEVICE_INVL:
+               val = DMA_CCMD_DEVICE_INVL|DMA_CCMD_DID(did)
+                       | DMA_CCMD_SID(source_id) | DMA_CCMD_FM(function_mask);
+               break;
+       default:
+               BUG();
+       }
+       val |= DMA_CCMD_ICC;
+
+       spin_lock_irqsave(&iommu->register_lock, flag);
+       dmar_writeq(iommu->reg + DMAR_CCMD_REG, val);
+
+       /* Make sure hardware complete it */
+       IOMMU_WAIT_OP(iommu, DMAR_CCMD_REG,
+               dmar_readq, (!(val & DMA_CCMD_ICC)), val);
+
+       spin_unlock_irqrestore(&iommu->register_lock, flag);
+
+       /* flush context entry will implictly flush write buffer */
+       return 0;
+}
+
+static int inline iommu_flush_context_global(struct intel_iommu *iommu,
+       int non_present_entry_flush)
+{
+       return __iommu_flush_context(iommu, 0, 0, 0, DMA_CCMD_GLOBAL_INVL,
+               non_present_entry_flush);
+}
+
+static int inline iommu_flush_context_domain(struct intel_iommu *iommu, u16 did,
+       int non_present_entry_flush)
+{
+       return __iommu_flush_context(iommu, did, 0, 0, DMA_CCMD_DOMAIN_INVL,
+               non_present_entry_flush);
+}
+
+static int inline iommu_flush_context_device(struct intel_iommu *iommu,
+       u16 did, u16 source_id, u8 function_mask, int non_present_entry_flush)
+{
+       return __iommu_flush_context(iommu, did, source_id, function_mask,
+               DMA_CCMD_DEVICE_INVL, non_present_entry_flush);
+}
+
+/* return value determine if we need a write buffer flush */
+static int __iommu_flush_iotlb(struct intel_iommu *iommu, u16 did,
+       u64 addr, unsigned int size_order, u64 type,
+       int non_present_entry_flush)
+{
+       int tlb_offset = ecap_iotlb_offset(iommu->ecap);
+       u64 val = 0, val_iva = 0;
+       unsigned long flag;
+
+       /*
+        * In the non-present entry flush case, if hardware doesn't cache
+        * non-present entry we do nothing and if hardware cache non-present
+        * entry, we flush entries of domain 0 (the domain id is used to cache
+        * any non-present entries)
+        */
+       if (non_present_entry_flush) {
+               if (!cap_caching_mode(iommu->cap))
+                       return 1;
+               else
+                       did = 0;
+       }
+
+       switch (type) {
+       case DMA_TLB_GLOBAL_FLUSH:
+               /* global flush doesn't need set IVA_REG */
+               val = DMA_TLB_GLOBAL_FLUSH|DMA_TLB_IVT;
+               break;
+       case DMA_TLB_DSI_FLUSH:
+               val = DMA_TLB_DSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
+               break;
+       case DMA_TLB_PSI_FLUSH:
+               val = DMA_TLB_PSI_FLUSH|DMA_TLB_IVT|DMA_TLB_DID(did);
+               /* Note: always flush non-leaf currently */
+               val_iva = size_order | addr;
+               break;
+       default:
+               BUG();
+       }
+       /* Note: set drain read/write */
+#if 0
+       /*
+        * This is probably to be super secure.. Looks like we can
+        * ignore it without any impact.
+        */
+       if (cap_read_drain(iommu->cap))
+               val |= DMA_TLB_READ_DRAIN;
+#endif
+       if (cap_write_drain(iommu->cap))
+               val |= DMA_TLB_WRITE_DRAIN;
+
+       spin_lock_irqsave(&iommu->register_lock, flag);
+       /* Note: Only uses first TLB reg currently */
+       if (val_iva)
+               dmar_writeq(iommu->reg + tlb_offset, val_iva);
+       dmar_writeq(iommu->reg + tlb_offset + 8, val);
+
+       /* Make sure hardware complete it */
+       IOMMU_WAIT_OP(iommu, tlb_offset + 8,
+               dmar_readq, (!(val & DMA_TLB_IVT)), val);
+
+       spin_unlock_irqrestore(&iommu->register_lock, flag);
+
+       /* check IOTLB invalidation granularity */
+       if (DMA_TLB_IAIG(val) == 0)
+               printk(KERN_ERR"IOMMU: flush IOTLB failed\n");
+       if (DMA_TLB_IAIG(val) != DMA_TLB_IIRG(type))
+               pr_debug("IOMMU: tlb flush request %Lx, actual %Lx\n",
+                       DMA_TLB_IIRG(type), DMA_TLB_IAIG(val));
+       /* flush context entry will implictly flush write buffer */
+       return 0;
+}
+
+static int inline iommu_flush_iotlb_global(struct intel_iommu *iommu,
+       int non_present_entry_flush)
+{
+       return __iommu_flush_iotlb(iommu, 0, 0, 0, DMA_TLB_GLOBAL_FLUSH,
+               non_present_entry_flush);
+}
+
+static int inline iommu_flush_iotlb_dsi(struct intel_iommu *iommu, u16 did,
+       int non_present_entry_flush)
+{
+       return __iommu_flush_iotlb(iommu, did, 0, 0, DMA_TLB_DSI_FLUSH,
+               non_present_entry_flush);
+}
+
+static int iommu_get_alignment(u64 base, unsigned int size)
+{
+       int t = 0;
+       u64 end;
+
+       end = base + size - 1;
+       while (base != end) {
+               t++;
+               base >>= 1;
+               end >>= 1;
+       }
+       return t;
+}
+
+static int iommu_flush_iotlb_psi(struct intel_iommu *iommu, u16 did,
+       u64 addr, unsigned int pages, int non_present_entry_flush)
+{
+       unsigned int align;
+
+       BUG_ON(addr & (~PAGE_MASK_4K));
+       BUG_ON(pages == 0);
+
+       /* Fallback to domain selective flush if no PSI support */
+       if (!cap_pgsel_inv(iommu->cap))
+               return iommu_flush_iotlb_dsi(iommu, did,
+                       non_present_entry_flush);
+
+       /*
+        * PSI requires page size to be 2 ^ x, and the base address is naturally
+        * aligned to the size
+        */
+       align = iommu_get_alignment(addr >> PAGE_SHIFT_4K, pages);
+       /* Fallback to domain selective flush if size is too big */
+       if (align > cap_max_amask_val(iommu->cap))
+               return iommu_flush_iotlb_dsi(iommu, did,
+                       non_present_entry_flush);
+
+       addr >>= PAGE_SHIFT_4K + align;
+       addr <<= PAGE_SHIFT_4K + align;
+
+       return __iommu_flush_iotlb(iommu, did, addr, align,
+               DMA_TLB_PSI_FLUSH, non_present_entry_flush);
+}
+
+static int iommu_enable_translation(struct intel_iommu *iommu)
+{
+       u32 sts;
+       unsigned long flags;
+
+       spin_lock_irqsave(&iommu->register_lock, flags);
+       writel(iommu->gcmd|DMA_GCMD_TE, iommu->reg + DMAR_GCMD_REG);
+
+       /* Make sure hardware complete it */
+       IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
+               readl, (sts & DMA_GSTS_TES), sts);
+
+       iommu->gcmd |= DMA_GCMD_TE;
+       spin_unlock_irqrestore(&iommu->register_lock, flags);
+       return 0;
+}
+
+static int iommu_disable_translation(struct intel_iommu *iommu)
+{
+       u32 sts;
+       unsigned long flag;
+
+       spin_lock_irqsave(&iommu->register_lock, flag);
+       iommu->gcmd &= ~DMA_GCMD_TE;
+       writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
+
+       /* Make sure hardware complete it */
+       IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG,
+               readl, (!(sts & DMA_GSTS_TES)), sts);
+
+       spin_unlock_irqrestore(&iommu->register_lock, flag);
+       return 0;
+}
+
+static int iommu_init_domains(struct intel_iommu *iommu)
+{
+       unsigned long ndomains;
+       unsigned long nlongs;
+
+       ndomains = cap_ndoms(iommu->cap);
+       pr_debug("Number of Domains supportd <%ld>\n", ndomains);
+       nlongs = BITS_TO_LONGS(ndomains);
+
+       /* TBD: there might be 64K domains,
+        * consider other allocation for future chip
+        */
+       iommu->domain_ids = kcalloc(nlongs, sizeof(unsigned long), GFP_KERNEL);
+       if (!iommu->domain_ids) {
+               printk(KERN_ERR "Allocating domain id array failed\n");
+               return -ENOMEM;
+       }
+       iommu->domains = kcalloc(ndomains, sizeof(struct dmar_domain *),
+                       GFP_KERNEL);
+       if (!iommu->domains) {
+               printk(KERN_ERR "Allocating domain array failed\n");
+               kfree(iommu->domain_ids);
+               return -ENOMEM;
+       }
+
+       /*
+        * if Caching mode is set, then invalid translations are tagged
+        * with domainid 0. Hence we need to pre-allocate it.
+        */
+       if (cap_caching_mode(iommu->cap))
+               set_bit(0, iommu->domain_ids);
+       return 0;
+}
+
+static struct intel_iommu *alloc_iommu(struct dmar_drhd_unit *drhd)
+{
+       struct intel_iommu *iommu;
+       int ret;
+       int map_size;
+       u32 ver;
+
+       iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
+       if (!iommu)
+               return NULL;
+       iommu->reg = ioremap(drhd->reg_base_addr, PAGE_SIZE_4K);
+       if (!iommu->reg) {
+               printk(KERN_ERR "IOMMU: can't map the region\n");
+               goto error;
+       }
+       iommu->cap = dmar_readq(iommu->reg + DMAR_CAP_REG);
+       iommu->ecap = dmar_readq(iommu->reg + DMAR_ECAP_REG);
+
+       /* the registers might be more than one page */
+       map_size = max_t(int, ecap_max_iotlb_offset(iommu->ecap),
+               cap_max_fault_reg_offset(iommu->cap));
+       map_size = PAGE_ALIGN_4K(map_size);
+       if (map_size > PAGE_SIZE_4K) {
+               iounmap(iommu->reg);
+               iommu->reg = ioremap(drhd->reg_base_addr, map_size);
+               if (!iommu->reg) {
+                       printk(KERN_ERR "IOMMU: can't map the region\n");
+                       goto error;
+               }
+       }
+
+       ver = readl(iommu->reg + DMAR_VER_REG);
+       pr_debug("IOMMU %llx: ver %d:%d cap %llx ecap %llx\n",
+               drhd->reg_base_addr, DMAR_VER_MAJOR(ver), DMAR_VER_MINOR(ver),
+               iommu->cap, iommu->ecap);
+       ret = iommu_init_domains(iommu);
+       if (ret)
+               goto error_unmap;
+       spin_lock_init(&iommu->lock);
+       spin_lock_init(&iommu->register_lock);
+
+       drhd->iommu = iommu;
+       return iommu;
+error_unmap:
+       iounmap(iommu->reg);
+       iommu->reg = 0;
+error:
+       kfree(iommu);
+       return NULL;
+}
+
+static void domain_exit(struct dmar_domain *domain);
+static void free_iommu(struct intel_iommu *iommu)
+{
+       struct dmar_domain *domain;
+       int i;
+
+       if (!iommu)
+               return;
+
+       i = find_first_bit(iommu->domain_ids, cap_ndoms(iommu->cap));
+       for (; i < cap_ndoms(iommu->cap); ) {
+               domain = iommu->domains[i];
+               clear_bit(i, iommu->domain_ids);
+               domain_exit(domain);
+               i = find_next_bit(iommu->domain_ids,
+                       cap_ndoms(iommu->cap), i+1);
+       }
+
+       if (iommu->gcmd & DMA_GCMD_TE)
+               iommu_disable_translation(iommu);
+
+       if (iommu->irq) {
+               set_irq_data(iommu->irq, NULL);
+               /* This will mask the irq */
+               free_irq(iommu->irq, iommu);
+               destroy_irq(iommu->irq);
+       }
+
+       kfree(iommu->domains);
+       kfree(iommu->domain_ids);
+
+       /* free context mapping */
+       free_context_table(iommu);
+
+       if (iommu->reg)
+               iounmap(iommu->reg);
+       kfree(iommu);
+}
+
+static struct dmar_domain * iommu_alloc_domain(struct intel_iommu *iommu)
+{
+       unsigned long num;
+       unsigned long ndomains;
+       struct dmar_domain *domain;
+       unsigned long flags;
+
+       domain = alloc_domain_mem();
+       if (!domain)
+               return NULL;
+
+       ndomains = cap_ndoms(iommu->cap);
+
+       spin_lock_irqsave(&iommu->lock, flags);
+       num = find_first_zero_bit(iommu->domain_ids, ndomains);
+       if (num >= ndomains) {
+               spin_unlock_irqrestore(&iommu->lock, flags);
+               free_domain_mem(domain);
+               printk(KERN_ERR "IOMMU: no free domain ids\n");
+               return NULL;
+       }
+
+       set_bit(num, iommu->domain_ids);
+       domain->id = num;
+       domain->iommu = iommu;
+       iommu->domains[num] = domain;
+       spin_unlock_irqrestore(&iommu->lock, flags);
+
+       return domain;
+}
+
+static void iommu_free_domain(struct dmar_domain *domain)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&domain->iommu->lock, flags);
+       clear_bit(domain->id, domain->iommu->domain_ids);
+       spin_unlock_irqrestore(&domain->iommu->lock, flags);
+}
+
+static struct iova_domain reserved_iova_list;
+
+static void dmar_init_reserved_ranges(void)
+{
+       struct pci_dev *pdev = NULL;
+       struct iova *iova;
+       int i;
+       u64 addr, size;
+
+       init_iova_domain(&reserved_iova_list);
+
+       /* IOAPIC ranges shouldn't be accessed by DMA */
+       iova = reserve_iova(&reserved_iova_list, IOVA_PFN(IOAPIC_RANGE_START),
+               IOVA_PFN(IOAPIC_RANGE_END));
+       if (!iova)
+               printk(KERN_ERR "Reserve IOAPIC range failed\n");
+
+       /* Reserve all PCI MMIO to avoid peer-to-peer access */
+       for_each_pci_dev(pdev) {
+               struct resource *r;
+
+               for (i = 0; i < PCI_NUM_RESOURCES; i++) {
+                       r = &pdev->resource[i];
+                       if (!r->flags || !(r->flags & IORESOURCE_MEM))
+                               continue;
+                       addr = r->start;
+                       addr &= PAGE_MASK_4K;
+                       size = r->end - addr;
+                       size = PAGE_ALIGN_4K(size);
+                       iova = reserve_iova(&reserved_iova_list, IOVA_PFN(addr),
+                               IOVA_PFN(size + addr) - 1);
+                       if (!iova)
+                               printk(KERN_ERR "Reserve iova failed\n");
+               }
+       }
+
+}
+
+static void domain_reserve_special_ranges(struct dmar_domain *domain)
+{
+       copy_reserved_iova(&reserved_iova_list, &domain->iovad);
+}
+
+static inline int guestwidth_to_adjustwidth(int gaw)
+{
+       int agaw;
+       int r = (gaw - 12) % 9;
+
+       if (r == 0)
+               agaw = gaw;
+       else
+               agaw = gaw + 9 - r;
+       if (agaw > 64)
+               agaw = 64;
+       return agaw;
+}
+
+static int domain_init(struct dmar_domain *domain, int guest_width)
+{
+       struct intel_iommu *iommu;
+       int adjust_width, agaw;
+       unsigned long sagaw;
+
+       init_iova_domain(&domain->iovad);
+       spin_lock_init(&domain->mapping_lock);
+
+       domain_reserve_special_ranges(domain);
+
+       /* calculate AGAW */
+       iommu = domain->iommu;
+       if (guest_width > cap_mgaw(iommu->cap))
+               guest_width = cap_mgaw(iommu->cap);
+       domain->gaw = guest_width;
+       adjust_width = guestwidth_to_adjustwidth(guest_width);
+       agaw = width_to_agaw(adjust_width);
+       sagaw = cap_sagaw(iommu->cap);
+       if (!test_bit(agaw, &sagaw)) {
+               /* hardware doesn't support it, choose a bigger one */
+               pr_debug("IOMMU: hardware doesn't support agaw %d\n", agaw);
+               agaw = find_next_bit(&sagaw, 5, agaw);
+               if (agaw >= 5)
+                       return -ENODEV;
+       }
+       domain->agaw = agaw;
+       INIT_LIST_HEAD(&domain->devices);
+
+       /* always allocate the top pgd */
+       domain->pgd = (struct dma_pte *)alloc_pgtable_page();
+       if (!domain->pgd)
+               return -ENOMEM;
+       __iommu_flush_cache(iommu, domain->pgd, PAGE_SIZE_4K);
+       return 0;
+}
+
+static void domain_exit(struct dmar_domain *domain)
+{
+       u64 end;
+
+       /* Domain 0 is reserved, so dont process it */
+       if (!domain)
+               return;
+
+       domain_remove_dev_info(domain);
+       /* destroy iovas */
+       put_iova_domain(&domain->iovad);
+       end = DOMAIN_MAX_ADDR(domain->gaw);
+       end = end & (~PAGE_MASK_4K);
+
+       /* clear ptes */
+       dma_pte_clear_range(domain, 0, end);
+
+       /* free page tables */
+       dma_pte_free_pagetable(domain, 0, end);
+
+       iommu_free_domain(domain);
+       free_domain_mem(domain);
+}
+
+static int domain_context_mapping_one(struct dmar_domain *domain,
+               u8 bus, u8 devfn)
+{
+       struct context_entry *context;
+       struct intel_iommu *iommu = domain->iommu;
+       unsigned long flags;
+
+       pr_debug("Set context mapping for %02x:%02x.%d\n",
+               bus, PCI_SLOT(devfn), PCI_FUNC(devfn));
+       BUG_ON(!domain->pgd);
+       context = device_to_context_entry(iommu, bus, devfn);
+       if (!context)
+               return -ENOMEM;
+       spin_lock_irqsave(&iommu->lock, flags);
+       if (context_present(*context)) {
+               spin_unlock_irqrestore(&iommu->lock, flags);
+               return 0;
+       }
+
+       context_set_domain_id(*context, domain->id);
+       context_set_address_width(*context, domain->agaw);
+       context_set_address_root(*context, virt_to_phys(domain->pgd));
+       context_set_translation_type(*context, CONTEXT_TT_MULTI_LEVEL);
+       context_set_fault_enable(*context);
+       context_set_present(*context);
+       __iommu_flush_cache(iommu, context, sizeof(*context));
+
+       /* it's a non-present to present mapping */
+       if (iommu_flush_context_device(iommu, domain->id,
+                       (((u16)bus) << 8) | devfn, DMA_CCMD_MASK_NOBIT, 1))
+               iommu_flush_write_buffer(iommu);
+       else
+               iommu_flush_iotlb_dsi(iommu, 0, 0);
+       spin_unlock_irqrestore(&iommu->lock, flags);
+       return 0;
+}
+
+static int
+domain_context_mapping(struct dmar_domain *domain, struct pci_dev *pdev)
+{
+       int ret;
+       struct pci_dev *tmp, *parent;
+
+       ret = domain_context_mapping_one(domain, pdev->bus->number,
+               pdev->devfn);
+       if (ret)
+               return ret;
+
+       /* dependent device mapping */
+       tmp = pci_find_upstream_pcie_bridge(pdev);
+       if (!tmp)
+               return 0;
+       /* Secondary interface's bus number and devfn 0 */
+       parent = pdev->bus->self;
+       while (parent != tmp) {
+               ret = domain_context_mapping_one(domain, parent->bus->number,
+                       parent->devfn);
+               if (ret)
+                       return ret;
+               parent = parent->bus->self;
+       }
+       if (tmp->is_pcie) /* this is a PCIE-to-PCI bridge */
+               return domain_context_mapping_one(domain,
+                       tmp->subordinate->number, 0);
+       else /* this is a legacy PCI bridge */
+               return domain_context_mapping_one(domain,
+                       tmp->bus->number, tmp->devfn);
+}
+
+static int domain_context_mapped(struct dmar_domain *domain,
+       struct pci_dev *pdev)
+{
+       int ret;
+       struct pci_dev *tmp, *parent;
+
+       ret = device_context_mapped(domain->iommu,
+               pdev->bus->number, pdev->devfn);
+       if (!ret)
+               return ret;
+       /* dependent device mapping */
+       tmp = pci_find_upstream_pcie_bridge(pdev);
+       if (!tmp)
+               return ret;
+       /* Secondary interface's bus number and devfn 0 */
+       parent = pdev->bus->self;
+       while (parent != tmp) {
+               ret = device_context_mapped(domain->iommu, parent->bus->number,
+                       parent->devfn);
+               if (!ret)
+                       return ret;
+               parent = parent->bus->self;
+       }
+       if (tmp->is_pcie)
+               return device_context_mapped(domain->iommu,
+                       tmp->subordinate->number, 0);
+       else
+               return device_context_mapped(domain->iommu,
+                       tmp->bus->number, tmp->devfn);
+}
+
+static int
+domain_page_mapping(struct dmar_domain *domain, dma_addr_t iova,
+                       u64 hpa, size_t size, int prot)
+{
+       u64 start_pfn, end_pfn;
+       struct dma_pte *pte;
+       int index;
+
+       if ((prot & (DMA_PTE_READ|DMA_PTE_WRITE)) == 0)
+               return -EINVAL;
+       iova &= PAGE_MASK_4K;
+       start_pfn = ((u64)hpa) >> PAGE_SHIFT_4K;
+       end_pfn = (PAGE_ALIGN_4K(((u64)hpa) + size)) >> PAGE_SHIFT_4K;
+       index = 0;
+       while (start_pfn < end_pfn) {
+               pte = addr_to_dma_pte(domain, iova + PAGE_SIZE_4K * index);
+               if (!pte)
+                       return -ENOMEM;
+               /* We don't need lock here, nobody else
+                * touches the iova range
+                */
+               BUG_ON(dma_pte_addr(*pte));
+               dma_set_pte_addr(*pte, start_pfn << PAGE_SHIFT_4K);
+               dma_set_pte_prot(*pte, prot);
+               __iommu_flush_cache(domain->iommu, pte, sizeof(*pte));
+               start_pfn++;
+               index++;
+       }
+       return 0;
+}
+
+static void detach_domain_for_dev(struct dmar_domain *domain, u8 bus, u8 devfn)
+{
+       clear_context_table(domain->iommu, bus, devfn);
+       iommu_flush_context_global(domain->iommu, 0);
+       iommu_flush_iotlb_global(domain->iommu, 0);
+}
+
+static void domain_remove_dev_info(struct dmar_domain *domain)
+{
+       struct device_domain_info *info;
+       unsigned long flags;
+
+       spin_lock_irqsave(&device_domain_lock, flags);
+       while (!list_empty(&domain->devices)) {
+               info = list_entry(domain->devices.next,
+                       struct device_domain_info, link);
+               list_del(&info->link);
+               list_del(&info->global);
+               if (info->dev)
+                       info->dev->sysdata = NULL;
+               spin_unlock_irqrestore(&device_domain_lock, flags);
+
+               detach_domain_for_dev(info->domain, info->bus, info->devfn);
+               free_devinfo_mem(info);
+
+               spin_lock_irqsave(&device_domain_lock, flags);
+       }
+       spin_unlock_irqrestore(&device_domain_lock, flags);
+}
+
+/*
+ * find_domain
+ * Note: we use struct pci_dev->sysdata stores the info
+ */
+struct dmar_domain *
+find_domain(struct pci_dev *pdev)
+{
+       struct device_domain_info *info;
+
+       /* No lock here, assumes no domain exit in normal case */
+       info = pdev->sysdata;
+       if (info)
+               return info->domain;
+       return NULL;
+}
+
+static int dmar_pci_device_match(struct pci_dev *devices[], int cnt,
+     struct pci_dev *dev)
+{
+       int index;
+
+       while (dev) {
+               for (index = 0; index < cnt; index ++)
+                       if (dev == devices[index])
+                               return 1;
+
+               /* Check our parent */
+               dev = dev->bus->self;
+       }
+
+       return 0;
+}
+
+static struct dmar_drhd_unit *
+dmar_find_matched_drhd_unit(struct pci_dev *dev)
+{
+       struct dmar_drhd_unit *drhd = NULL;
+
+       list_for_each_entry(drhd, &dmar_drhd_units, list) {
+               if (drhd->include_all || dmar_pci_device_match(drhd->devices,
+                                               drhd->devices_cnt, dev))
+                       return drhd;
+       }
+
+       return NULL;
+}
+
+/* domain is initialized */
+static struct dmar_domain *get_domain_for_dev(struct pci_dev *pdev, int gaw)
+{
+       struct dmar_domain *domain, *found = NULL;
+       struct intel_iommu *iommu;
+       struct dmar_drhd_unit *drhd;
+       struct device_domain_info *info, *tmp;
+       struct pci_dev *dev_tmp;
+       unsigned long flags;
+       int bus = 0, devfn = 0;
+
+       domain = find_domain(pdev);
+       if (domain)
+               return domain;
+
+       dev_tmp = pci_find_upstream_pcie_bridge(pdev);
+       if (dev_tmp) {
+               if (dev_tmp->is_pcie) {
+                       bus = dev_tmp->subordinate->number;
+                       devfn = 0;
+               } else {
+                       bus = dev_tmp->bus->number;
+                       devfn = dev_tmp->devfn;
+               }
+               spin_lock_irqsave(&device_domain_lock, flags);
+               list_for_each_entry(info, &device_domain_list, global) {
+                       if (info->bus == bus && info->devfn == devfn) {
+                               found = info->domain;
+                               break;
+                       }
+               }
+               spin_unlock_irqrestore(&device_domain_lock, flags);
+               /* pcie-pci bridge already has a domain, uses it */
+               if (found) {
+                       domain = found;
+                       goto found_domain;
+               }
+       }
+
+       /* Allocate new domain for the device */
+       drhd = dmar_find_matched_drhd_unit(pdev);
+       if (!drhd) {
+               printk(KERN_ERR "IOMMU: can't find DMAR for device %s\n",
+                       pci_name(pdev));
+               return NULL;
+       }
+       iommu = drhd->iommu;
+
+       domain = iommu_alloc_domain(iommu);
+       if (!domain)
+               goto error;
+
+       if (domain_init(domain, gaw)) {
+               domain_exit(domain);
+               goto error;
+       }
+
+       /* register pcie-to-pci device */
+       if (dev_tmp) {
+               info = alloc_devinfo_mem();
+               if (!info) {
+                       domain_exit(domain);
+                       goto error;
+               }
+               info->bus = bus;
+               info->devfn = devfn;
+               info->dev = NULL;
+               info->domain = domain;
+               /* This domain is shared by devices under p2p bridge */
+               domain->flags |= DOMAIN_FLAG_MULTIPLE_DEVICES;
+
+               /* pcie-to-pci bridge already has a domain, uses it */
+               found = NULL;
+               spin_lock_irqsave(&device_domain_lock, flags);
+               list_for_each_entry(tmp, &device_domain_list, global) {
+                       if (tmp->bus == bus && tmp->devfn == devfn) {
+                               found = tmp->domain;
+                               break;
+                       }
+               }
+               if (found) {
+                       free_devinfo_mem(info);
+                       domain_exit(domain);
+                       domain = found;
+               } else {
+                       list_add(&info->link, &domain->devices);
+                       list_add(&info->global, &device_domain_list);
+               }
+               spin_unlock_irqrestore(&device_domain_lock, flags);
+       }
+
+found_domain:
+       info = alloc_devinfo_mem();
+       if (!info)
+               goto error;
+       info->bus = pdev->bus->number;
+       info->devfn = pdev->devfn;
+       info->dev = pdev;
+       info->domain = domain;
+       spin_lock_irqsave(&device_domain_lock, flags);
+       /* somebody is fast */
+       found = find_domain(pdev);
+       if (found != NULL) {
+               spin_unlock_irqrestore(&device_domain_lock, flags);
+               if (found != domain) {
+                       domain_exit(domain);
+                       domain = found;
+               }
+               free_devinfo_mem(info);
+               return domain;
+       }
+       list_add(&info->link, &domain->devices);
+       list_add(&info->global, &device_domain_list);
+       pdev->sysdata = info;
+       spin_unlock_irqrestore(&device_domain_lock, flags);
+       return domain;
+error:
+       /* recheck it here, maybe others set it */
+       return find_domain(pdev);
+}
+
+static int iommu_prepare_identity_map(struct pci_dev *pdev, u64 start, u64 end)
+{
+       struct dmar_domain *domain;
+       unsigned long size;
+       u64 base;
+       int ret;
+
+       printk(KERN_INFO
+               "IOMMU: Setting identity map for device %s [0x%Lx - 0x%Lx]\n",
+               pci_name(pdev), start, end);
+       /* page table init */
+       domain = get_domain_for_dev(pdev, DEFAULT_DOMAIN_ADDRESS_WIDTH);
+       if (!domain)
+               return -ENOMEM;
+
+       /* The address might not be aligned */
+       base = start & PAGE_MASK_4K;
+       size = end - base;
+       size = PAGE_ALIGN_4K(size);
+       if (!reserve_iova(&domain->iovad, IOVA_PFN(base),
+                       IOVA_PFN(base + size) - 1)) {
+               printk(KERN_ERR "IOMMU: reserve iova failed\n");
+               ret = -ENOMEM;
+               goto error;
+       }
+
+       pr_debug("Mapping reserved region %lx@%llx for %s\n",
+               size, base, pci_name(pdev));
+       /*
+        * RMRR range might have overlap with physical memory range,
+        * clear it first
+        */
+       dma_pte_clear_range(domain, base, base + size);
+
+       ret = domain_page_mapping(domain, base, base, size,
+               DMA_PTE_READ|DMA_PTE_WRITE);
+       if (ret)
+               goto error;
+
+       /* context entry init */
+       ret = domain_context_mapping(domain, pdev);
+       if (!ret)
+               return 0;
+error:
+       domain_exit(domain);
+       return ret;
+
+}
+
+static inline int iommu_prepare_rmrr_dev(struct dmar_rmrr_unit *rmrr,
+       struct pci_dev *pdev)
+{
+       if (pdev->sysdata == DUMMY_DEVICE_DOMAIN_INFO)
+               return 0;
+       return iommu_prepare_identity_map(pdev, rmrr->base_address,
+               rmrr->end_address + 1);
+}
+
+int __init init_dmars(void)
+{
+       struct dmar_drhd_unit *drhd;
+       struct dmar_rmrr_unit *rmrr;
+       struct pci_dev *pdev;
+       struct intel_iommu *iommu;
+       int ret, unit = 0;
+
+       /*
+        * for each drhd
+        *    allocate root
+        *    initialize and program root entry to not present
+        * endfor
+        */
+       for_each_drhd_unit(drhd) {
+               if (drhd->ignored)
+                       continue;
+               iommu = alloc_iommu(drhd);
+               if (!iommu) {
+                       ret = -ENOMEM;
+                       goto error;
+               }
+
+               /*
+                * TBD:
+                * we could share the same root & context tables
+                * amoung all IOMMU's. Need to Split it later.
+                */
+               ret = iommu_alloc_root_entry(iommu);
+               if (ret) {
+                       printk(KERN_ERR "IOMMU: allocate root entry failed\n");
+                       goto error;
+               }
+       }
+
+       /*
+        * For each rmrr
+        *   for each dev attached to rmrr
+        *   do
+        *     locate drhd for dev, alloc domain for dev
+        *     allocate free domain
+        *     allocate page table entries for rmrr
+        *     if context not allocated for bus
+        *           allocate and init context
+        *           set present in root table for this bus
+        *     init context with domain, translation etc
+        *    endfor
+        * endfor
+        */
+       for_each_rmrr_units(rmrr) {
+               int i;
+               for (i = 0; i < rmrr->devices_cnt; i++) {
+                       pdev = rmrr->devices[i];
+                       /* some BIOS lists non-exist devices in DMAR table */
+                       if (!pdev)
+                               continue;
+                       ret = iommu_prepare_rmrr_dev(rmrr, pdev);
+                       if (ret)
+                               printk(KERN_ERR
+                                "IOMMU: mapping reserved region failed\n");
+               }
+       }
+
+       /*
+        * for each drhd
+        *   enable fault log
+        *   global invalidate context cache
+        *   global invalidate iotlb
+        *   enable translation
+        */
+       for_each_drhd_unit(drhd) {
+               if (drhd->ignored)
+                       continue;
+               iommu = drhd->iommu;
+               sprintf (iommu->name, "dmar%d", unit++);
+
+               iommu_flush_write_buffer(iommu);
+
+               iommu_set_root_entry(iommu);
+
+               iommu_flush_context_global(iommu, 0);
+               iommu_flush_iotlb_global(iommu, 0);
+
+               ret = iommu_enable_translation(iommu);
+               if (ret)
+                       goto error;
+       }
+
+       return 0;
+error:
+       for_each_drhd_unit(drhd) {
+               if (drhd->ignored)
+                       continue;
+               iommu = drhd->iommu;
+               free_iommu(iommu);
+       }
+       return ret;
+}
+
+static inline u64 aligned_size(u64 host_addr, size_t size)
+{
+       u64 addr;
+       addr = (host_addr & (~PAGE_MASK_4K)) + size;
+       return PAGE_ALIGN_4K(addr);
+}
+
+struct iova *
+iommu_alloc_iova(struct dmar_domain *domain, void *host_addr, size_t size,
+               u64 start, u64 end)
+{
+       u64 start_addr;
+       struct iova *piova;
+
+       /* Make sure it's in range */
+       if ((start > DOMAIN_MAX_ADDR(domain->gaw)) || end < start)
+               return NULL;
+
+       end = min_t(u64, DOMAIN_MAX_ADDR(domain->gaw), end);
+       start_addr = PAGE_ALIGN_4K(start);
+       size = aligned_size((u64)host_addr, size);
+       if (!size || (start_addr + size > end))
+               return NULL;
+
+       piova = alloc_iova(&domain->iovad,
+                       size >> PAGE_SHIFT_4K, IOVA_PFN(end));
+
+       return piova;
+}
+
+static dma_addr_t __intel_map_single(struct device *dev, void *addr,
+       size_t size, int dir, u64 *flush_addr, unsigned int *flush_size)
+{
+       struct dmar_domain *domain;
+       struct pci_dev *pdev = to_pci_dev(dev);
+       int ret;
+       int prot = 0;
+       struct iova *iova = NULL;
+       u64 start_addr;
+
+       addr = (void *)virt_to_phys(addr);
+
+       domain = get_domain_for_dev(pdev,
+                       DEFAULT_DOMAIN_ADDRESS_WIDTH);
+       if (!domain) {
+               printk(KERN_ERR
+                       "Allocating domain for %s failed", pci_name(pdev));
+               return 0;
+       }
+
+       start_addr = IOVA_START_ADDR;
+
+       if (pdev->dma_mask <= DMA_32BIT_MASK) {
+               iova = iommu_alloc_iova(domain, addr, size, start_addr,
+                       pdev->dma_mask);
+       } else  {
+               /*
+                * First try to allocate an io virtual address in
+                * DMA_32BIT_MASK and if that fails then try allocating
+                * from higer range
+                */
+               iova = iommu_alloc_iova(domain, addr, size, start_addr,
+                       DMA_32BIT_MASK);
+               if (!iova)
+                       iova = iommu_alloc_iova(domain, addr, size, start_addr,
+                       pdev->dma_mask);
+       }
+
+       if (!iova) {
+               printk(KERN_ERR"Allocating iova for %s failed", pci_name(pdev));
+               return 0;
+       }
+
+       /* make sure context mapping is ok */
+       if (unlikely(!domain_context_mapped(domain, pdev))) {
+               ret = domain_context_mapping(domain, pdev);
+               if (ret)
+                       goto error;
+       }
+
+       /*
+        * Check if DMAR supports zero-length reads on write only
+        * mappings..
+        */
+       if (dir == DMA_TO_DEVICE || dir == DMA_BIDIRECTIONAL || \
+                       !cap_zlr(domain->iommu->cap))
+               prot |= DMA_PTE_READ;
+       if (dir == DMA_FROM_DEVICE || dir == DMA_BIDIRECTIONAL)
+               prot |= DMA_PTE_WRITE;
+       /*
+        * addr - (addr + size) might be partial page, we should map the whole
+        * page.  Note: if two part of one page are separately mapped, we
+        * might have two guest_addr mapping to the same host addr, but this
+        * is not a big problem
+        */
+       ret = domain_page_mapping(domain, iova->pfn_lo << PAGE_SHIFT_4K,
+               ((u64)addr) & PAGE_MASK_4K,
+               (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT_4K, prot);
+       if (ret)
+               goto error;
+
+       pr_debug("Device %s request: %lx@%llx mapping: %lx@%llx, dir %d\n",
+               pci_name(pdev), size, (u64)addr,
+               (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT_4K,
+               (u64)(iova->pfn_lo << PAGE_SHIFT_4K), dir);
+
+       *flush_addr = iova->pfn_lo << PAGE_SHIFT_4K;
+       *flush_size = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT_4K;
+       return (iova->pfn_lo << PAGE_SHIFT_4K) + ((u64)addr & (~PAGE_MASK_4K));
+error:
+       __free_iova(&domain->iovad, iova);
+       printk(KERN_ERR"Device %s request: %lx@%llx dir %d --- failed\n",
+               pci_name(pdev), size, (u64)addr, dir);
+       return 0;
+}
+
+static dma_addr_t intel_map_single(struct device *hwdev, void *addr,
+       size_t size, int dir)
+{
+       struct pci_dev *pdev = to_pci_dev(hwdev);
+       dma_addr_t ret;
+       struct dmar_domain *domain;
+       u64 flush_addr;
+       unsigned int flush_size;
+
+       BUG_ON(dir == DMA_NONE);
+       if (pdev->sysdata == DUMMY_DEVICE_DOMAIN_INFO)
+               return virt_to_bus(addr);
+
+       ret = __intel_map_single(hwdev, addr, size,
+                       dir, &flush_addr, &flush_size);
+       if (ret) {
+               domain = find_domain(pdev);
+               /* it's a non-present to present mapping */
+               if (iommu_flush_iotlb_psi(domain->iommu, domain->id,
+                               flush_addr, flush_size >> PAGE_SHIFT_4K, 1))
+                       iommu_flush_write_buffer(domain->iommu);
+       }
+       return ret;
+}
+
+static void __intel_unmap_single(struct device *dev, dma_addr_t dev_addr,
+       size_t size, int dir, u64 *flush_addr, unsigned int *flush_size)
+{
+       struct dmar_domain *domain;
+       struct pci_dev *pdev = to_pci_dev(dev);
+       struct iova *iova;
+
+       domain = find_domain(pdev);
+       BUG_ON(!domain);
+
+       iova = find_iova(&domain->iovad, IOVA_PFN(dev_addr));
+       if (!iova) {
+               *flush_size = 0;
+               return;
+       }
+       pr_debug("Device %s unmapping: %lx@%llx\n",
+               pci_name(pdev),
+               (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT_4K,
+               (u64)(iova->pfn_lo << PAGE_SHIFT_4K));
+
+       *flush_addr = iova->pfn_lo << PAGE_SHIFT_4K;
+       *flush_size = (iova->pfn_hi - iova->pfn_lo + 1) << PAGE_SHIFT_4K;
+       /*  clear the whole page, not just dev_addr - (dev_addr + size) */
+       dma_pte_clear_range(domain, *flush_addr, *flush_addr + *flush_size);
+       /* free page tables */
+       dma_pte_free_pagetable(domain, *flush_addr, *flush_addr + *flush_size);
+       /* free iova */
+       __free_iova(&domain->iovad, iova);
+}
+
+static void intel_unmap_single(struct device *dev, dma_addr_t dev_addr,
+       size_t size, int dir)
+{
+       struct pci_dev *pdev = to_pci_dev(dev);
+       struct dmar_domain *domain;
+       u64 flush_addr;
+       unsigned int flush_size;
+
+       if (pdev->sysdata == DUMMY_DEVICE_DOMAIN_INFO)
+               return;
+
+       domain = find_domain(pdev);
+       __intel_unmap_single(dev, dev_addr, size,
+               dir, &flush_addr, &flush_size);
+       if (flush_size == 0)
+               return;
+       if (iommu_flush_iotlb_psi(domain->iommu, domain->id, flush_addr,
+                       flush_size >> PAGE_SHIFT_4K, 0))
+               iommu_flush_write_buffer(domain->iommu);
+}
+
+static void * intel_alloc_coherent(struct device *hwdev, size_t size,
+                      dma_addr_t *dma_handle, gfp_t flags)
+{
+       void *vaddr;
+       int order;
+
+       size = PAGE_ALIGN_4K(size);
+       order = get_order(size);
+       flags &= ~(GFP_DMA | GFP_DMA32);
+
+       vaddr = (void *)__get_free_pages(flags, order);
+       if (!vaddr)
+               return NULL;
+       memset(vaddr, 0, size);
+
+       *dma_handle = intel_map_single(hwdev, vaddr, size, DMA_BIDIRECTIONAL);
+       if (*dma_handle)
+               return vaddr;
+       free_pages((unsigned long)vaddr, order);
+       return NULL;
+}
+
+static void intel_free_coherent(struct device *hwdev, size_t size,
+       void *vaddr, dma_addr_t dma_handle)
+{
+       int order;
+
+       size = PAGE_ALIGN_4K(size);
+       order = get_order(size);
+
+       intel_unmap_single(hwdev, dma_handle, size, DMA_BIDIRECTIONAL);
+       free_pages((unsigned long)vaddr, order);
+}
+
+static void intel_unmap_sg(struct device *hwdev, struct scatterlist *sg,
+       int nelems, int dir)
+{
+       int i;
+       struct pci_dev *pdev = to_pci_dev(hwdev);
+       struct dmar_domain *domain;
+       u64 flush_addr;
+       unsigned int flush_size;
+
+       if (pdev->sysdata == DUMMY_DEVICE_DOMAIN_INFO)
+               return;
+
+       domain = find_domain(pdev);
+       for (i = 0; i < nelems; i++, sg++)
+               __intel_unmap_single(hwdev, sg->dma_address,
+                       sg->dma_length, dir, &flush_addr, &flush_size);
+
+       if (iommu_flush_iotlb_dsi(domain->iommu, domain->id, 0))
+               iommu_flush_write_buffer(domain->iommu);
+}
+
+#define SG_ENT_VIRT_ADDRESS(sg)        (page_address((sg)->page) + (sg)->offset)
+static int intel_nontranslate_map_sg(struct device *hddev,
+       struct scatterlist *sg, int nelems, int dir)
+{
+       int i;
+
+       for (i = 0; i < nelems; i++) {
+               struct scatterlist *s = &sg[i];
+               BUG_ON(!s->page);
+               s->dma_address = virt_to_bus(SG_ENT_VIRT_ADDRESS(s));
+               s->dma_length = s->length;
+       }
+       return nelems;
+}
+
+static int intel_map_sg(struct device *hwdev, struct scatterlist *sg,
+       int nelems, int dir)
+{
+       void *addr;
+       int i;
+       dma_addr_t dma_handle;
+       struct pci_dev *pdev = to_pci_dev(hwdev);
+       struct dmar_domain *domain;
+       u64 flush_addr;
+       unsigned int flush_size;
+
+       BUG_ON(dir == DMA_NONE);
+       if (pdev->sysdata == DUMMY_DEVICE_DOMAIN_INFO)
+               return intel_nontranslate_map_sg(hwdev, sg, nelems, dir);
+
+       for (i = 0; i < nelems; i++, sg++) {
+               addr = SG_ENT_VIRT_ADDRESS(sg);
+               dma_handle = __intel_map_single(hwdev, addr,
+                               sg->length, dir, &flush_addr, &flush_size);
+               if (!dma_handle) {
+                       intel_unmap_sg(hwdev, sg - i, i, dir);
+                       sg[0].dma_length = 0;
+                       return 0;
+               }
+               sg->dma_address = dma_handle;
+               sg->dma_length = sg->length;
+       }
+
+       domain = find_domain(pdev);
+
+       /* it's a non-present to present mapping */
+       if (iommu_flush_iotlb_dsi(domain->iommu, domain->id, 1))
+               iommu_flush_write_buffer(domain->iommu);
+       return nelems;
+}
+
+static struct dma_mapping_ops intel_dma_ops = {
+       .alloc_coherent = intel_alloc_coherent,
+       .free_coherent = intel_free_coherent,
+       .map_single = intel_map_single,
+       .unmap_single = intel_unmap_single,
+       .map_sg = intel_map_sg,
+       .unmap_sg = intel_unmap_sg,
+};
+
+static inline int iommu_domain_cache_init(void)
+{
+       int ret = 0;
+
+       iommu_domain_cache = kmem_cache_create("iommu_domain",
+                                        sizeof(struct dmar_domain),
+                                        0,
+                                        SLAB_HWCACHE_ALIGN,
+
+                                        NULL);
+       if (!iommu_domain_cache) {
+               printk(KERN_ERR "Couldn't create iommu_domain cache\n");
+               ret = -ENOMEM;
+       }
+
+       return ret;
+}
+
+static inline int iommu_devinfo_cache_init(void)
+{
+       int ret = 0;
+
+       iommu_devinfo_cache = kmem_cache_create("iommu_devinfo",
+                                        sizeof(struct device_domain_info),
+                                        0,
+                                        SLAB_HWCACHE_ALIGN,
+
+                                        NULL);
+       if (!iommu_devinfo_cache) {
+               printk(KERN_ERR "Couldn't create devinfo cache\n");
+               ret = -ENOMEM;
+       }
+
+       return ret;
+}
+
+static inline int iommu_iova_cache_init(void)
+{
+       int ret = 0;
+
+       iommu_iova_cache = kmem_cache_create("iommu_iova",
+                                        sizeof(struct iova),
+                                        0,
+                                        SLAB_HWCACHE_ALIGN,
+
+                                        NULL);
+       if (!iommu_iova_cache) {
+               printk(KERN_ERR "Couldn't create iova cache\n");
+               ret = -ENOMEM;
+       }
+
+       return ret;
+}
+
+static int __init iommu_init_mempool(void)
+{
+       int ret;
+       ret = iommu_iova_cache_init();
+       if (ret)
+               return ret;
+
+       ret = iommu_domain_cache_init();
+       if (ret)
+               goto domain_error;
+
+       ret = iommu_devinfo_cache_init();
+       if (!ret)
+               return ret;
+
+       kmem_cache_destroy(iommu_domain_cache);
+domain_error:
+       kmem_cache_destroy(iommu_iova_cache);
+
+       return -ENOMEM;
+}
+
+static void __init iommu_exit_mempool(void)
+{
+       kmem_cache_destroy(iommu_devinfo_cache);
+       kmem_cache_destroy(iommu_domain_cache);
+       kmem_cache_destroy(iommu_iova_cache);
+
+}
+
+void __init detect_intel_iommu(void)
+{
+       if (swiotlb || no_iommu || iommu_detected || dmar_disabled)
+               return;
+       if (early_dmar_detect()) {
+               iommu_detected = 1;
+       }
+}
+
+static void __init init_no_remapping_devices(void)
+{
+       struct dmar_drhd_unit *drhd;
+
+       for_each_drhd_unit(drhd) {
+               if (!drhd->include_all) {
+                       int i;
+                       for (i = 0; i < drhd->devices_cnt; i++)
+                               if (drhd->devices[i] != NULL)
+                                       break;
+                       /* ignore DMAR unit if no pci devices exist */
+                       if (i == drhd->devices_cnt)
+                               drhd->ignored = 1;
+               }
+       }
+
+       if (dmar_map_gfx)
+               return;
+
+       for_each_drhd_unit(drhd) {
+               int i;
+               if (drhd->ignored || drhd->include_all)
+                       continue;
+
+               for (i = 0; i < drhd->devices_cnt; i++)
+                       if (drhd->devices[i] &&
+                               !IS_GFX_DEVICE(drhd->devices[i]))
+                               break;
+
+               if (i < drhd->devices_cnt)
+                       continue;
+
+               /* bypass IOMMU if it is just for gfx devices */
+               drhd->ignored = 1;
+               for (i = 0; i < drhd->devices_cnt; i++) {
+                       if (!drhd->devices[i])
+                               continue;
+                       drhd->devices[i]->sysdata = DUMMY_DEVICE_DOMAIN_INFO;
+               }
+       }
+}
+
+int __init intel_iommu_init(void)
+{
+       int ret = 0;
+
+       if (no_iommu || swiotlb || dmar_disabled)
+               return -ENODEV;
+
+       if (dmar_table_init())
+               return  -ENODEV;
+
+       iommu_init_mempool();
+       dmar_init_reserved_ranges();
+
+       init_no_remapping_devices();
+
+       ret = init_dmars();
+       if (ret) {
+               printk(KERN_ERR "IOMMU: dmar init failed\n");
+               put_iova_domain(&reserved_iova_list);
+               iommu_exit_mempool();
+               return ret;
+       }
+       printk(KERN_INFO
+       "PCI-DMA: Intel(R) Virtualization Technology for Directed I/O\n");
+
+       force_iommu = 1;
+       dma_ops = &intel_dma_ops;
+       return 0;
+}
diff --git a/drivers/pci/intel-iommu.h b/drivers/pci/intel-iommu.h
new file mode 100644 (file)
index 0000000..71dda6b
--- /dev/null
@@ -0,0 +1,318 @@
+/*
+ * Copyright (c) 2006, Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+ * Place - Suite 330, Boston, MA 02111-1307 USA.
+ *
+ * Copyright (C) Ashok Raj <ashok.raj@intel.com>
+ * Copyright (C) Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
+ */
+
+#ifndef _INTEL_IOMMU_H_
+#define _INTEL_IOMMU_H_
+
+#include <linux/types.h>
+#include <linux/msi.h>
+#include "iova.h"
+#include <linux/io.h>
+
+/*
+ * Intel IOMMU register specification per version 1.0 public spec.
+ */
+
+#define        DMAR_VER_REG    0x0     /* Arch version supported by this IOMMU */
+#define        DMAR_CAP_REG    0x8     /* Hardware supported capabilities */
+#define        DMAR_ECAP_REG   0x10    /* Extended capabilities supported */
+#define        DMAR_GCMD_REG   0x18    /* Global command register */
+#define        DMAR_GSTS_REG   0x1c    /* Global status register */
+#define        DMAR_RTADDR_REG 0x20    /* Root entry table */
+#define        DMAR_CCMD_REG   0x28    /* Context command reg */
+#define        DMAR_FSTS_REG   0x34    /* Fault Status register */
+#define        DMAR_FECTL_REG  0x38    /* Fault control register */
+#define        DMAR_FEDATA_REG 0x3c    /* Fault event interrupt data register */
+#define        DMAR_FEADDR_REG 0x40    /* Fault event interrupt addr register */
+#define        DMAR_FEUADDR_REG 0x44   /* Upper address register */
+#define        DMAR_AFLOG_REG  0x58    /* Advanced Fault control */
+#define        DMAR_PMEN_REG   0x64    /* Enable Protected Memory Region */
+#define        DMAR_PLMBASE_REG 0x68   /* PMRR Low addr */
+#define        DMAR_PLMLIMIT_REG 0x6c  /* PMRR low limit */
+#define        DMAR_PHMBASE_REG 0x70   /* pmrr high base addr */
+#define        DMAR_PHMLIMIT_REG 0x78  /* pmrr high limit */
+
+#define OFFSET_STRIDE          (9)
+/*
+#define dmar_readl(dmar, reg) readl(dmar + reg)
+#define dmar_readq(dmar, reg) ({ \
+               u32 lo, hi; \
+               lo = readl(dmar + reg); \
+               hi = readl(dmar + reg + 4); \
+               (((u64) hi) << 32) + lo; })
+*/
+static inline u64 dmar_readq(void *addr)
+{
+       u32 lo, hi;
+       lo = readl(addr);
+       hi = readl(addr + 4);
+       return (((u64) hi) << 32) + lo;
+}
+
+static inline void dmar_writeq(void __iomem *addr, u64 val)
+{
+       writel((u32)val, addr);
+       writel((u32)(val >> 32), addr + 4);
+}
+
+#define DMAR_VER_MAJOR(v)              (((v) & 0xf0) >> 4)
+#define DMAR_VER_MINOR(v)              ((v) & 0x0f)
+
+/*
+ * Decoding Capability Register
+ */
+#define cap_read_drain(c)      (((c) >> 55) & 1)
+#define cap_write_drain(c)     (((c) >> 54) & 1)
+#define cap_max_amask_val(c)   (((c) >> 48) & 0x3f)
+#define cap_num_fault_regs(c)  ((((c) >> 40) & 0xff) + 1)
+#define cap_pgsel_inv(c)       (((c) >> 39) & 1)
+
+#define cap_super_page_val(c)  (((c) >> 34) & 0xf)
+#define cap_super_offset(c)    (((find_first_bit(&cap_super_page_val(c), 4)) \
+                                       * OFFSET_STRIDE) + 21)
+
+#define cap_fault_reg_offset(c)        ((((c) >> 24) & 0x3ff) * 16)
+#define cap_max_fault_reg_offset(c) \
+       (cap_fault_reg_offset(c) + cap_num_fault_regs(c) * 16)
+
+#define cap_zlr(c)             (((c) >> 22) & 1)
+#define cap_isoch(c)           (((c) >> 23) & 1)
+#define cap_mgaw(c)            ((((c) >> 16) & 0x3f) + 1)
+#define cap_sagaw(c)           (((c) >> 8) & 0x1f)
+#define cap_caching_mode(c)    (((c) >> 7) & 1)
+#define cap_phmr(c)            (((c) >> 6) & 1)
+#define cap_plmr(c)            (((c) >> 5) & 1)
+#define cap_rwbf(c)            (((c) >> 4) & 1)
+#define cap_afl(c)             (((c) >> 3) & 1)
+#define cap_ndoms(c)           (((unsigned long)1) << (4 + 2 * ((c) & 0x7)))
+/*
+ * Extended Capability Register
+ */
+
+#define ecap_niotlb_iunits(e)  ((((e) >> 24) & 0xff) + 1)
+#define ecap_iotlb_offset(e)   ((((e) >> 8) & 0x3ff) * 16)
+#define ecap_max_iotlb_offset(e) \
+       (ecap_iotlb_offset(e) + ecap_niotlb_iunits(e) * 16)
+#define ecap_coherent(e)       ((e) & 0x1)
+
+
+/* IOTLB_REG */
+#define DMA_TLB_GLOBAL_FLUSH (((u64)1) << 60)
+#define DMA_TLB_DSI_FLUSH (((u64)2) << 60)
+#define DMA_TLB_PSI_FLUSH (((u64)3) << 60)
+#define DMA_TLB_IIRG(type) ((type >> 60) & 7)
+#define DMA_TLB_IAIG(val) (((val) >> 57) & 7)
+#define DMA_TLB_READ_DRAIN (((u64)1) << 49)
+#define DMA_TLB_WRITE_DRAIN (((u64)1) << 48)
+#define DMA_TLB_DID(id)        (((u64)((id) & 0xffff)) << 32)
+#define DMA_TLB_IVT (((u64)1) << 63)
+#define DMA_TLB_IH_NONLEAF (((u64)1) << 6)
+#define DMA_TLB_MAX_SIZE (0x3f)
+
+/* GCMD_REG */
+#define DMA_GCMD_TE (((u32)1) << 31)
+#define DMA_GCMD_SRTP (((u32)1) << 30)
+#define DMA_GCMD_SFL (((u32)1) << 29)
+#define DMA_GCMD_EAFL (((u32)1) << 28)
+#define DMA_GCMD_WBF (((u32)1) << 27)
+
+/* GSTS_REG */
+#define DMA_GSTS_TES (((u32)1) << 31)
+#define DMA_GSTS_RTPS (((u32)1) << 30)
+#define DMA_GSTS_FLS (((u32)1) << 29)
+#define DMA_GSTS_AFLS (((u32)1) << 28)
+#define DMA_GSTS_WBFS (((u32)1) << 27)
+
+/* CCMD_REG */
+#define DMA_CCMD_ICC (((u64)1) << 63)
+#define DMA_CCMD_GLOBAL_INVL (((u64)1) << 61)
+#define DMA_CCMD_DOMAIN_INVL (((u64)2) << 61)
+#define DMA_CCMD_DEVICE_INVL (((u64)3) << 61)
+#define DMA_CCMD_FM(m) (((u64)((m) & 0x3)) << 32)
+#define DMA_CCMD_MASK_NOBIT 0
+#define DMA_CCMD_MASK_1BIT 1
+#define DMA_CCMD_MASK_2BIT 2
+#define DMA_CCMD_MASK_3BIT 3
+#define DMA_CCMD_SID(s) (((u64)((s) & 0xffff)) << 16)
+#define DMA_CCMD_DID(d) ((u64)((d) & 0xffff))
+
+/* FECTL_REG */
+#define DMA_FECTL_IM (((u32)1) << 31)
+
+/* FSTS_REG */
+#define DMA_FSTS_PPF ((u32)2)
+#define DMA_FSTS_PFO ((u32)1)
+#define dma_fsts_fault_record_index(s) (((s) >> 8) & 0xff)
+
+/* FRCD_REG, 32 bits access */
+#define DMA_FRCD_F (((u32)1) << 31)
+#define dma_frcd_type(d) ((d >> 30) & 1)
+#define dma_frcd_fault_reason(c) (c & 0xff)
+#define dma_frcd_source_id(c) (c & 0xffff)
+#define dma_frcd_page_addr(d) (d & (((u64)-1) << 12)) /* low 64 bit */
+
+/*
+ * 0: Present
+ * 1-11: Reserved
+ * 12-63: Context Ptr (12 - (haw-1))
+ * 64-127: Reserved
+ */
+struct root_entry {
+       u64     val;
+       u64     rsvd1;
+};
+#define ROOT_ENTRY_NR (PAGE_SIZE_4K/sizeof(struct root_entry))
+static inline bool root_present(struct root_entry *root)
+{
+       return (root->val & 1);
+}
+static inline void set_root_present(struct root_entry *root)
+{
+       root->val |= 1;
+}
+static inline void set_root_value(struct root_entry *root, unsigned long value)
+{
+       root->val |= value & PAGE_MASK_4K;
+}
+
+struct context_entry;
+static inline struct context_entry *
+get_context_addr_from_root(struct root_entry *root)
+{
+       return (struct context_entry *)
+               (root_present(root)?phys_to_virt(
+               root->val & PAGE_MASK_4K):
+               NULL);
+}
+
+/*
+ * low 64 bits:
+ * 0: present
+ * 1: fault processing disable
+ * 2-3: translation type
+ * 12-63: address space root
+ * high 64 bits:
+ * 0-2: address width
+ * 3-6: aval
+ * 8-23: domain id
+ */
+struct context_entry {
+       u64 lo;
+       u64 hi;
+};
+#define context_present(c) ((c).lo & 1)
+#define context_fault_disable(c) (((c).lo >> 1) & 1)
+#define context_translation_type(c) (((c).lo >> 2) & 3)
+#define context_address_root(c) ((c).lo & PAGE_MASK_4K)
+#define context_address_width(c) ((c).hi &  7)
+#define context_domain_id(c) (((c).hi >> 8) & ((1 << 16) - 1))
+
+#define context_set_present(c) do {(c).lo |= 1;} while (0)
+#define context_set_fault_enable(c) \
+       do {(c).lo &= (((u64)-1) << 2) | 1;} while (0)
+#define context_set_translation_type(c, val) \
+       do { \
+               (c).lo &= (((u64)-1) << 4) | 3; \
+               (c).lo |= ((val) & 3) << 2; \
+       } while (0)
+#define CONTEXT_TT_MULTI_LEVEL 0
+#define context_set_address_root(c, val) \
+       do {(c).lo |= (val) & PAGE_MASK_4K;} while (0)
+#define context_set_address_width(c, val) do {(c).hi |= (val) & 7;} while (0)
+#define context_set_domain_id(c, val) \
+       do {(c).hi |= ((val) & ((1 << 16) - 1)) << 8;} while (0)
+#define context_clear_entry(c) do {(c).lo = 0; (c).hi = 0;} while (0)
+
+/*
+ * 0: readable
+ * 1: writable
+ * 2-6: reserved
+ * 7: super page
+ * 8-11: available
+ * 12-63: Host physcial address
+ */
+struct dma_pte {
+       u64 val;
+};
+#define dma_clear_pte(p)       do {(p).val = 0;} while (0)
+
+#define DMA_PTE_READ (1)
+#define DMA_PTE_WRITE (2)
+
+#define dma_set_pte_readable(p) do {(p).val |= DMA_PTE_READ;} while (0)
+#define dma_set_pte_writable(p) do {(p).val |= DMA_PTE_WRITE;} while (0)
+#define dma_set_pte_prot(p, prot) \
+               do {(p).val = ((p).val & ~3) | ((prot) & 3); } while (0)
+#define dma_pte_addr(p) ((p).val & PAGE_MASK_4K)
+#define dma_set_pte_addr(p, addr) do {\
+               (p).val |= ((addr) & PAGE_MASK_4K); } while (0)
+#define dma_pte_present(p) (((p).val & 3) != 0)
+
+struct intel_iommu;
+
+struct dmar_domain {
+       int     id;                     /* domain id */
+       struct intel_iommu *iommu;      /* back pointer to owning iommu */
+
+       struct list_head devices;       /* all devices' list */
+       struct iova_domain iovad;       /* iova's that belong to this domain */
+
+       struct dma_pte  *pgd;           /* virtual address */
+       spinlock_t      mapping_lock;   /* page table lock */
+       int             gaw;            /* max guest address width */
+
+       /* adjusted guest address width, 0 is level 2 30-bit */
+       int             agaw;
+
+#define DOMAIN_FLAG_MULTIPLE_DEVICES 1
+       int             flags;
+};
+
+/* PCI domain-device relationship */
+struct device_domain_info {
+       struct list_head link;  /* link to domain siblings */
+       struct list_head global; /* link to global list */
+       u8 bus;                 /* PCI bus numer */
+       u8 devfn;               /* PCI devfn number */
+       struct pci_dev *dev; /* it's NULL for PCIE-to-PCI bridge */
+       struct dmar_domain *domain; /* pointer to domain */
+};
+
+extern int init_dmars(void);
+
+struct intel_iommu {
+       void __iomem    *reg; /* Pointer to hardware regs, virtual addr */
+       u64             cap;
+       u64             ecap;
+       unsigned long   *domain_ids; /* bitmap of domains */
+       struct dmar_domain **domains; /* ptr to domains */
+       int             seg;
+       u32             gcmd; /* Holds TE, EAFL. Don't need SRTP, SFL, WBF */
+       spinlock_t      lock; /* protect context, domain ids */
+       spinlock_t      register_lock; /* protect register handling */
+       struct root_entry *root_entry; /* virtual address */
+
+       unsigned int irq;
+       unsigned char name[7];    /* Device Name */
+       struct msi_msg saved_msg;
+       struct sys_device sysdev;
+};
+
+#endif
index 8d3e0e3..7d683dc 100644 (file)
 
 #include <linux/acpi.h>
 #include <linux/types.h>
+#include <linux/msi.h>
 
+#ifdef CONFIG_DMAR
+struct intel_iommu;
+
+/* Intel IOMMU detection and initialization functions */
+extern void detect_intel_iommu(void);
+extern int intel_iommu_init(void);
 
 extern int dmar_table_init(void);
 extern int early_dmar_detect(void);
@@ -49,4 +56,19 @@ struct dmar_rmrr_unit {
        int     devices_cnt;            /* target device count */
 };
 
+#define for_each_drhd_unit(drhd) \
+       list_for_each_entry(drhd, &dmar_drhd_units, list)
+#define for_each_rmrr_units(rmrr) \
+       list_for_each_entry(rmrr, &dmar_rmrr_units, list)
+#else
+static inline void detect_intel_iommu(void)
+{
+       return;
+}
+static inline int intel_iommu_init(void)
+{
+       return -ENODEV;
+}
+
+#endif /* !CONFIG_DMAR */
 #endif /* __DMAR_H__ */