[PATCH] /dev/mem: validate mmap requests
authorBjorn Helgaas <bjorn.helgaas@hp.com>
Sun, 8 Jan 2006 09:04:13 +0000 (01:04 -0800)
committerLinus Torvalds <torvalds@g5.osdl.org>
Mon, 9 Jan 2006 04:14:02 +0000 (20:14 -0800)
Add a hook so architectures can validate /dev/mem mmap requests.

This is analogous to validation we already perform in the read/write
paths.

The identity mapping scheme used on ia64 requires that each 16MB or
64MB granule be accessed with exactly one attribute (write-back or
uncacheable).  This avoids "attribute aliasing", which can cause a
machine check.

Sample problem scenario:
  - Machine supports VGA, so it has uncacheable (UC) MMIO at 640K-768K
  - efi_memmap_init() discards any write-back (WB) memory in the first granule
  - Application (e.g., "hwinfo") mmaps /dev/mem, offset 0
  - hwinfo receives UC mapping (the default, since memmap says "no WB here")
  - Machine check abort (on chipsets that don't support UC access to WB
    memory, e.g., sx1000)

In the scenario above, the only choices are
  - Use WB for hwinfo mmap.  Can't do this because it causes attribute
    aliasing with the UC mapping for the VGA MMIO space.
  - Use UC for hwinfo mmap.  Can't do this because the chipset may not
    support UC for that region.
  - Disallow the hwinfo mmap with -EINVAL.  That's what this patch does.

Signed-off-by: Bjorn Helgaas <bjorn.helgaas@hp.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
arch/ia64/kernel/efi.c
drivers/char/mem.c
include/asm-ia64/io.h

index a3aa45cbcfa03e460dab2d8696259126688407fb..c485a3b32ba8ba4a31ce380635441ab4e3ff1729 100644 (file)
@@ -247,6 +247,32 @@ typedef struct kern_memdesc {
 
 static kern_memdesc_t *kern_memmap;
 
+#define efi_md_size(md)        (md->num_pages << EFI_PAGE_SHIFT)
+
+static inline u64
+kmd_end(kern_memdesc_t *kmd)
+{
+       return (kmd->start + (kmd->num_pages << EFI_PAGE_SHIFT));
+}
+
+static inline u64
+efi_md_end(efi_memory_desc_t *md)
+{
+       return (md->phys_addr + efi_md_size(md));
+}
+
+static inline int
+efi_wb(efi_memory_desc_t *md)
+{
+       return (md->attribute & EFI_MEMORY_WB);
+}
+
+static inline int
+efi_uc(efi_memory_desc_t *md)
+{
+       return (md->attribute & EFI_MEMORY_UC);
+}
+
 static void
 walk (efi_freemem_callback_t callback, void *arg, u64 attr)
 {
@@ -595,8 +621,8 @@ efi_get_iobase (void)
        return 0;
 }
 
-u32
-efi_mem_type (unsigned long phys_addr)
+static efi_memory_desc_t *
+efi_memory_descriptor (unsigned long phys_addr)
 {
        void *efi_map_start, *efi_map_end, *p;
        efi_memory_desc_t *md;
@@ -610,13 +636,13 @@ efi_mem_type (unsigned long phys_addr)
                md = p;
 
                if (phys_addr - md->phys_addr < (md->num_pages << EFI_PAGE_SHIFT))
-                        return md->type;
+                        return md;
        }
        return 0;
 }
 
-u64
-efi_mem_attributes (unsigned long phys_addr)
+static int
+efi_memmap_has_mmio (void)
 {
        void *efi_map_start, *efi_map_end, *p;
        efi_memory_desc_t *md;
@@ -629,36 +655,98 @@ efi_mem_attributes (unsigned long phys_addr)
        for (p = efi_map_start; p < efi_map_end; p += efi_desc_size) {
                md = p;
 
-               if (phys_addr - md->phys_addr < (md->num_pages << EFI_PAGE_SHIFT))
-                       return md->attribute;
+               if (md->type == EFI_MEMORY_MAPPED_IO)
+                       return 1;
        }
        return 0;
 }
+
+u32
+efi_mem_type (unsigned long phys_addr)
+{
+       efi_memory_desc_t *md = efi_memory_descriptor(phys_addr);
+
+       if (md)
+               return md->type;
+       return 0;
+}
+
+u64
+efi_mem_attributes (unsigned long phys_addr)
+{
+       efi_memory_desc_t *md = efi_memory_descriptor(phys_addr);
+
+       if (md)
+               return md->attribute;
+       return 0;
+}
 EXPORT_SYMBOL(efi_mem_attributes);
 
+/*
+ * Determines whether the memory at phys_addr supports the desired
+ * attribute (WB, UC, etc).  If this returns 1, the caller can safely
+ * access *size bytes at phys_addr with the specified attribute.
+ */
+static int
+efi_mem_attribute_range (unsigned long phys_addr, unsigned long *size, u64 attr)
+{
+       efi_memory_desc_t *md = efi_memory_descriptor(phys_addr);
+       unsigned long md_end;
+
+       if (!md || (md->attribute & attr) != attr)
+               return 0;
+
+       do {
+               md_end = efi_md_end(md);
+               if (phys_addr + *size <= md_end)
+                       return 1;
+
+               md = efi_memory_descriptor(md_end);
+               if (!md || (md->attribute & attr) != attr) {
+                       *size = md_end - phys_addr;
+                       return 1;
+               }
+       } while (md);
+       return 0;
+}
+
+/*
+ * For /dev/mem, we only allow read & write system calls to access
+ * write-back memory, because read & write don't allow the user to
+ * control access size.
+ */
 int
 valid_phys_addr_range (unsigned long phys_addr, unsigned long *size)
 {
-       void *efi_map_start, *efi_map_end, *p;
-       efi_memory_desc_t *md;
-       u64 efi_desc_size;
+       return efi_mem_attribute_range(phys_addr, size, EFI_MEMORY_WB);
+}
 
-       efi_map_start = __va(ia64_boot_param->efi_memmap);
-       efi_map_end   = efi_map_start + ia64_boot_param->efi_memmap_size;
-       efi_desc_size = ia64_boot_param->efi_memdesc_size;
+/*
+ * We allow mmap of anything in the EFI memory map that supports
+ * either write-back or uncacheable access.  For uncacheable regions,
+ * the supported access sizes are system-dependent, and the user is
+ * responsible for using the correct size.
+ *
+ * Note that this doesn't currently allow access to hot-added memory,
+ * because that doesn't appear in the boot-time EFI memory map.
+ */
+int
+valid_mmap_phys_addr_range (unsigned long phys_addr, unsigned long *size)
+{
+       if (efi_mem_attribute_range(phys_addr, size, EFI_MEMORY_WB))
+               return 1;
 
-       for (p = efi_map_start; p < efi_map_end; p += efi_desc_size) {
-               md = p;
+       if (efi_mem_attribute_range(phys_addr, size, EFI_MEMORY_UC))
+               return 1;
 
-               if (phys_addr - md->phys_addr < (md->num_pages << EFI_PAGE_SHIFT)) {
-                       if (!(md->attribute & EFI_MEMORY_WB))
-                               return 0;
+       /*
+        * Some firmware doesn't report MMIO regions in the EFI memory map.
+        * The Intel BigSur (a.k.a. HP i2000) has this problem.  In this
+        * case, we can't use the EFI memory map to validate mmap requests.
+        */
+       if (!efi_memmap_has_mmio())
+               return 1;
 
-                       if (*size > md->phys_addr + (md->num_pages << EFI_PAGE_SHIFT) - phys_addr)
-                               *size = md->phys_addr + (md->num_pages << EFI_PAGE_SHIFT) - phys_addr;
-                       return 1;
-               }
-       }
        return 0;
 }
 
@@ -707,32 +795,6 @@ efi_uart_console_only(void)
        return 0;
 }
 
-#define efi_md_size(md)        (md->num_pages << EFI_PAGE_SHIFT)
-
-static inline u64
-kmd_end(kern_memdesc_t *kmd)
-{
-       return (kmd->start + (kmd->num_pages << EFI_PAGE_SHIFT));
-}
-
-static inline u64
-efi_md_end(efi_memory_desc_t *md)
-{
-       return (md->phys_addr + efi_md_size(md));
-}
-
-static inline int
-efi_wb(efi_memory_desc_t *md)
-{
-       return (md->attribute & EFI_MEMORY_WB);
-}
-
-static inline int
-efi_uc(efi_memory_desc_t *md)
-{
-       return (md->attribute & EFI_MEMORY_UC);
-}
-
 /*
  * Look for the first granule aligned memory descriptor memory
  * that is big enough to hold EFI memory map. Make sure this
index ce3ff8641191524c4deebfc0b03917e917c34445..5b2d18035073311e0bf7b150d3cca0a75f7b48b7 100644 (file)
@@ -101,6 +101,11 @@ static inline int valid_phys_addr_range(unsigned long addr, size_t *count)
 
        return 1;
 }
+
+static inline int valid_mmap_phys_addr_range(unsigned long addr, size_t *size)
+{
+       return 1;
+}
 #endif
 
 /*
@@ -244,15 +249,20 @@ static pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn,
 
 static int mmap_mem(struct file * file, struct vm_area_struct * vma)
 {
+       size_t size = vma->vm_end - vma->vm_start;
+
+       if (!valid_mmap_phys_addr_range(vma->vm_pgoff << PAGE_SHIFT, &size))
+               return -EINVAL;
+
        vma->vm_page_prot = phys_mem_access_prot(file, vma->vm_pgoff,
-                                                vma->vm_end - vma->vm_start,
+                                                size,
                                                 vma->vm_page_prot);
 
        /* Remap-pfn-range will mark the range VM_IO and VM_RESERVED */
        if (remap_pfn_range(vma,
                            vma->vm_start,
                            vma->vm_pgoff,
-                           vma->vm_end-vma->vm_start,
+                           size,
                            vma->vm_page_prot))
                return -EAGAIN;
        return 0;
index cf772a67f858763719a615f03ea8b8d117501973..b64fdb9854941cfcd816f003b99d3e66235dc34d 100644 (file)
@@ -89,6 +89,7 @@ phys_to_virt (unsigned long address)
 
 #define ARCH_HAS_VALID_PHYS_ADDR_RANGE
 extern int valid_phys_addr_range (unsigned long addr, size_t *count); /* efi.c */
+extern int valid_mmap_phys_addr_range (unsigned long addr, size_t *count);
 
 /*
  * The following two macros are deprecated and scheduled for removal.