]> git.openfabrics.org - ~emulex/infiniband.git/commitdiff
[ARM] Add section support to ioremap
authorRussell King <rmk@dyn-67.arm.linux.org.uk>
Thu, 29 Jun 2006 19:17:15 +0000 (20:17 +0100)
committerRussell King <rmk+kernel@arm.linux.org.uk>
Thu, 29 Jun 2006 21:14:30 +0000 (22:14 +0100)
Allow section mappings to be setup using ioremap() and torn down
with iounmap().  This requires additional support in the MM
context switch to ensure that mappings are properly synchronised
when mapped in.

Based an original implementation by Deepak Saxena, reworked and
ARMv6 support added by rmk.

Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
arch/arm/mm/ioremap.c
include/asm-arm/memory.h
include/asm-arm/mmu.h
include/asm-arm/mmu_context.h

index 7691cfdba56778d7f97768b66c243e66eb735dc2..6aa13d59c85872d650dcd19d00e965c23eb9c748 100644 (file)
 
 #include <asm/cacheflush.h>
 #include <asm/io.h>
+#include <asm/mmu_context.h>
+#include <asm/pgalloc.h>
 #include <asm/tlbflush.h>
+#include <asm/sizes.h>
+
+/*
+ * Used by ioremap() and iounmap() code to mark section-mapped I/O regions
+ * in vm_struct->flags field.
+ */
+#define VM_ARM_SECTION_MAPPING 0x80000000
 
 static inline void
 remap_area_pte(pte_t * pte, unsigned long address, unsigned long size,
@@ -113,10 +122,120 @@ remap_area_pages(unsigned long start, unsigned long pfn,
                dir++;
        } while (address && (address < end));
 
-       flush_cache_vmap(start, end);
        return err;
 }
 
+
+void __check_kvm_seq(struct mm_struct *mm)
+{
+       unsigned int seq;
+
+       do {
+               seq = init_mm.context.kvm_seq;
+               memcpy(pgd_offset(mm, VMALLOC_START),
+                      pgd_offset_k(VMALLOC_START),
+                      sizeof(pgd_t) * (pgd_index(VMALLOC_END) -
+                                       pgd_index(VMALLOC_START)));
+               mm->context.kvm_seq = seq;
+       } while (seq != init_mm.context.kvm_seq);
+}
+
+#ifndef CONFIG_SMP
+/*
+ * Section support is unsafe on SMP - If you iounmap and ioremap a region,
+ * the other CPUs will not see this change until their next context switch.
+ * Meanwhile, (eg) if an interrupt comes in on one of those other CPUs
+ * which requires the new ioremap'd region to be referenced, the CPU will
+ * reference the _old_ region.
+ *
+ * Note that get_vm_area() allocates a guard 4K page, so we need to mask
+ * the size back to 1MB aligned or we will overflow in the loop below.
+ */
+static void unmap_area_sections(unsigned long virt, unsigned long size)
+{
+       unsigned long addr = virt, end = virt + (size & ~SZ_1M);
+       pgd_t *pgd;
+
+       flush_cache_vunmap(addr, end);
+       pgd = pgd_offset_k(addr);
+       do {
+               pmd_t pmd, *pmdp = pmd_offset(pgd, addr);
+
+               pmd = *pmdp;
+               if (!pmd_none(pmd)) {
+                       /*
+                        * Clear the PMD from the page table, and
+                        * increment the kvm sequence so others
+                        * notice this change.
+                        *
+                        * Note: this is still racy on SMP machines.
+                        */
+                       pmd_clear(pmdp);
+                       init_mm.context.kvm_seq++;
+
+                       /*
+                        * Free the page table, if there was one.
+                        */
+                       if ((pmd_val(pmd) & PMD_TYPE_MASK) == PMD_TYPE_TABLE)
+                               pte_free_kernel(pmd_page_kernel(pmd));
+               }
+
+               addr += PGDIR_SIZE;
+               pgd++;
+       } while (addr < end);
+
+       /*
+        * Ensure that the active_mm is up to date - we want to
+        * catch any use-after-iounmap cases.
+        */
+       if (current->active_mm->context.kvm_seq != init_mm.context.kvm_seq)
+               __check_kvm_seq(current->active_mm);
+
+       flush_tlb_kernel_range(virt, end);
+}
+
+static int
+remap_area_sections(unsigned long virt, unsigned long pfn,
+                   unsigned long size, unsigned long flags)
+{
+       unsigned long prot, addr = virt, end = virt + size;
+       pgd_t *pgd;
+
+       /*
+        * Remove and free any PTE-based mapping, and
+        * sync the current kernel mapping.
+        */
+       unmap_area_sections(virt, size);
+
+       prot = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_DOMAIN(DOMAIN_IO) |
+              (flags & (L_PTE_CACHEABLE | L_PTE_BUFFERABLE));
+
+       /*
+        * ARMv6 and above need XN set to prevent speculative prefetches
+        * hitting IO.
+        */
+       if (cpu_architecture() >= CPU_ARCH_ARMv6)
+               prot |= PMD_SECT_XN;
+
+       pgd = pgd_offset_k(addr);
+       do {
+               pmd_t *pmd = pmd_offset(pgd, addr);
+
+               pmd[0] = __pmd(__pfn_to_phys(pfn) | prot);
+               pfn += SZ_1M >> PAGE_SHIFT;
+               pmd[1] = __pmd(__pfn_to_phys(pfn) | prot);
+               pfn += SZ_1M >> PAGE_SHIFT;
+               flush_pmd_entry(pmd);
+
+               addr += PGDIR_SIZE;
+               pgd++;
+       } while (addr < end);
+
+       return 0;
+}
+#endif
+
+
 /*
  * Remap an arbitrary physical address space into the kernel virtual
  * address space. Needed when the kernel wants to access high addresses
@@ -133,6 +252,7 @@ void __iomem *
 __ioremap_pfn(unsigned long pfn, unsigned long offset, size_t size,
              unsigned long flags)
 {
+       int err;
        unsigned long addr;
        struct vm_struct * area;
 
@@ -140,11 +260,22 @@ __ioremap_pfn(unsigned long pfn, unsigned long offset, size_t size,
        if (!area)
                return NULL;
        addr = (unsigned long)area->addr;
-       if (remap_area_pages(addr, pfn, size, flags)) {
+
+#ifndef CONFIG_SMP
+       if (!((__pfn_to_phys(pfn) | size | addr) & ~PMD_MASK)) {
+               area->flags |= VM_ARM_SECTION_MAPPING;
+               err = remap_area_sections(addr, pfn, size, flags);
+       } else
+#endif
+               err = remap_area_pages(addr, pfn, size, flags);
+
+       if (err) {
                vunmap((void *)addr);
                return NULL;
        }
-       return (void __iomem *) (offset + (char *)addr);
+
+       flush_cache_vmap(addr, addr + size);
+       return (void __iomem *) (offset + addr);
 }
 EXPORT_SYMBOL(__ioremap_pfn);
 
@@ -173,6 +304,34 @@ EXPORT_SYMBOL(__ioremap);
 
 void __iounmap(void __iomem *addr)
 {
-       vunmap((void *)(PAGE_MASK & (unsigned long)addr));
+       struct vm_struct **p, *tmp;
+       unsigned int section_mapping = 0;
+
+       addr = (void __iomem *)(PAGE_MASK & (unsigned long)addr);
+
+       /*
+        * If this is a section based mapping we need to handle it
+        * specially as the VM subysystem does not know how to handle
+        * such a beast. We need the lock here b/c we need to clear
+        * all the mappings before the area can be reclaimed
+        * by someone else.
+        */
+       write_lock(&vmlist_lock);
+       for (p = &vmlist ; (tmp = *p) ; p = &tmp->next) {
+               if((tmp->flags & VM_IOREMAP) && (tmp->addr == addr)) {
+                       if (tmp->flags & VM_ARM_SECTION_MAPPING) {
+                               *p = tmp->next;
+                               unmap_area_sections((unsigned long)tmp->addr,
+                                                   tmp->size);
+                               kfree(tmp);
+                               section_mapping = 1;
+                       }
+                       break;
+               }
+       }
+       write_unlock(&vmlist_lock);
+
+       if (!section_mapping)
+               vunmap(addr);
 }
 EXPORT_SYMBOL(__iounmap);
index 94f973b704f1578f1f88e42dec61a8ac2670007d..176a4fb04989a63aa461fe624cbe902d8892ca73 100644 (file)
  */
 #define XIP_VIRT_ADDR(physaddr)  (MODULE_START + ((physaddr) & 0x000fffff))
 
+/*
+ * Allow 2MB-aligned ioremap pages
+ */
+#define IOREMAP_MAX_ORDER      21
+
 #else /* CONFIG_MMU */
 
 /*
index 23dde52e0945f31c24ddd9fc4ddc1b8e66a74163..fe2a23b5627bc41639aa8d0626c152149366b4ee 100644 (file)
@@ -7,6 +7,7 @@ typedef struct {
 #if __LINUX_ARM_ARCH__ >= 6
        unsigned int id;
 #endif
+       unsigned int kvm_seq;
 } mm_context_t;
 
 #if __LINUX_ARM_ARCH__ >= 6
index 9fadb01e030d15a4fd67dda4ac143ff52406be46..d1a65b1edcaab236f0a65706380f60e85d26fc04 100644 (file)
@@ -17,6 +17,8 @@
 #include <asm/cacheflush.h>
 #include <asm/proc-fns.h>
 
+void __check_kvm_seq(struct mm_struct *mm);
+
 #if __LINUX_ARM_ARCH__ >= 6
 
 /*
@@ -45,13 +47,21 @@ static inline void check_context(struct mm_struct *mm)
 {
        if (unlikely((mm->context.id ^ cpu_last_asid) >> ASID_BITS))
                __new_context(mm);
+
+       if (unlikely(mm->context.kvm_seq != init_mm.context.kvm_seq))
+               __check_kvm_seq(mm);
 }
 
 #define init_new_context(tsk,mm)       (__init_new_context(tsk,mm),0)
 
 #else
 
-#define check_context(mm)              do { } while (0)
+static inline void check_context(struct mm_struct *mm)
+{
+       if (unlikely(mm->context.kvm_seq != init_mm.context.kvm_seq))
+               __check_kvm_seq(mm);
+}
+
 #define init_new_context(tsk,mm)       0
 
 #endif