drivers/xen/char/mem.c

   1 /*
   2  *  Originally from linux/drivers/char/mem.c
   3  *
   4  *  Copyright (C) 1991, 1992  Linus Torvalds
   5  *
   6  *  Added devfs support.
   7  *    Jan-11-1998, C. Scott Ananian <cananian@alumni.princeton.edu>
   8  *  Shared /dev/zero mmapping support, Feb 2000, Kanoj Sarcar <kanoj@sgi.com>
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/init.h>
  13 #include <linux/fs.h>
  14 #include <linux/capability.h>
  15 #include <linux/ptrace.h>
  16 #include <asm/uaccess.h>
  17 #include <asm/io.h>
  18 #include <asm/hypervisor.h>
  19
  20 static inline unsigned long size_inside_page(unsigned long start,
  21                                              unsigned long size)
  22 {
  23         unsigned long sz;
  24
  25         sz = PAGE_SIZE - (start & (PAGE_SIZE - 1));
  26
  27         return min(sz, size);
  28 }
  29
  30 static inline int uncached_access(struct file *file)
  31 {
  32         if (file->f_flags & O_DSYNC)
  33                 return 1;
  34         /* Xen sets correct MTRR type on non-RAM for us. */
  35         return 0;
  36 }
  37
  38 static inline int range_is_allowed(unsigned long pfn, unsigned long size)
  39 {
  40 #ifdef CONFIG_STRICT_DEVMEM
  41         u64 from = ((u64)pfn) << PAGE_SHIFT;
  42         u64 to = from + size;
  43         u64 cursor = from;
  44
  45         while (cursor < to) {
  46                 if (!devmem_is_allowed(pfn)) {
  47                         printk(KERN_INFO
  48                 "Program %s tried to access /dev/mem between %Lx->%Lx.\n",
  49                                 current->comm, from, to);
  50                         return 0;
  51                 }
  52                 cursor += PAGE_SIZE;
  53                 pfn++;
  54         }
  55 #endif
  56         return 1;
  57 }
  58
  59 /*
  60  * This funcion reads the *physical* memory. The f_pos points directly to the
  61  * memory location.
  62  */
  63 static ssize_t read_mem(struct file *file, char __user *buf,
  64                         size_t count, loff_t *ppos)
  65 {
  66         unsigned long p = *ppos;
  67         ssize_t read = 0, sz;
  68         void __iomem *v;
  69
  70         while (count > 0) {
  71                 unsigned long remaining;
  72
  73                 sz = size_inside_page(p, count);
  74
  75                 if (!range_is_allowed(p >> PAGE_SHIFT, count))
  76                         return -EPERM;
  77
  78                 v = ioremap(p, sz);
  79                 if (IS_ERR(v) || v == NULL) {
  80                         /*
  81                          * Some programs (e.g., dmidecode) groove off into
  82                          * weird RAM areas where no tables can possibly exist
  83                          * (because Xen will have stomped on them!). These
  84                          * programs get rather upset if we let them know that
  85                          * Xen failed their access, so we fake out a read of
  86                          * all zeroes.
  87                          */
  88                         if (clear_user(buf, count))
  89                                 return -EFAULT;
  90                         read += count;
  91                         break;
  92                 }
  93
  94                 remaining = copy_to_user(buf, v, sz);
  95                 iounmap(v);
  96                 if (remaining)
  97                         return -EFAULT;
  98
  99                 buf += sz;
 100                 p += sz;
 101                 count -= sz;
 102                 read += sz;
 103         }
 104
 105         *ppos += read;
 106         return read;
 107 }
 108
 109 static ssize_t write_mem(struct file *file, const char __user *buf,
 110                          size_t count, loff_t *ppos)
 111 {
 112         unsigned long p = *ppos, ignored;
 113         ssize_t written = 0, sz;
 114         void __iomem *v;
 115
 116         while (count > 0) {
 117                 sz = size_inside_page(p, count);
 118
 119                 if (!range_is_allowed(p >> PAGE_SHIFT, sz))
 120                         return -EPERM;
 121
 122                 v = ioremap(p, sz);
 123                 if (v == NULL)
 124                         break;
 125                 if (IS_ERR(v)) {
 126                         if (written == 0)
 127                                 return PTR_ERR(v);
 128                         break;
 129                 }
 130
 131                 ignored = copy_from_user(v, buf, sz);
 132                 iounmap(v);
 133                 if (ignored) {
 134                         written += sz - ignored;
 135                         if (written)
 136                                 break;
 137                         return -EFAULT;
 138                 }
 139                 buf += sz;
 140                 p += sz;
 141                 count -= sz;
 142                 written += sz;
 143         }
 144
 145         *ppos += written;
 146         return written;
 147 }
 148
 149 #ifndef ARCH_HAS_DEV_MEM_MMAP_MEM
 150 static struct vm_operations_struct mmap_mem_ops = {
 151 #ifdef CONFIG_HAVE_IOREMAP_PROT
 152         .access = generic_access_phys
 153 #endif
 154 };
 155
 156 static int xen_mmap_mem(struct file *file, struct vm_area_struct *vma)
 157 {
 158         size_t size = vma->vm_end - vma->vm_start;
 159
 160         if (uncached_access(file))
 161                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 162
 163         if (!range_is_allowed(vma->vm_pgoff, size))
 164                 return -EPERM;
 165
 166         if (!phys_mem_access_prot_allowed(file, vma->vm_pgoff, size,
 167                                                 &vma->vm_page_prot))
 168                 return -EINVAL;
 169
 170         vma->vm_ops = &mmap_mem_ops;
 171
 172         /* We want to return the real error code, not EAGAIN. */
 173         return direct_remap_pfn_range(vma, vma->vm_start, vma->vm_pgoff,
 174                                       size, vma->vm_page_prot, DOMID_IO);
 175 }
 176 #endif
 177
 178 /*
 179  * The memory devices use the full 32/64 bits of the offset, and so we cannot
 180  * check against negative addresses: they are ok. The return value is weird,
 181  * though, in that case (0).
 182  *
 183  * also note that seeking relative to the "end of file" isn't supported:
 184  * it has no meaning, so it returns -EINVAL.
 185  */
 186 static loff_t memory_lseek(struct file *file, loff_t offset, int orig)
 187 {
 188         loff_t ret;
 189
 190         mutex_lock(&file->f_path.dentry->d_inode->i_mutex);
 191         switch (orig) {
 192         case SEEK_CUR:
 193                 offset += file->f_pos;
 194         case SEEK_SET:
 195                 /* to avoid userland mistaking f_pos=-9 as -EBADF=-9 */
 196                 if ((unsigned long long)offset >= ~0xFFFULL) {
 197                         ret = -EOVERFLOW;
 198                         break;
 199                 }
 200                 file->f_pos = offset;
 201                 ret = file->f_pos;
 202                 force_successful_syscall_return();
 203                 break;
 204         default:
 205                 ret = -EINVAL;
 206         }
 207         mutex_unlock(&file->f_path.dentry->d_inode->i_mutex);
 208         return ret;
 209 }
 210
 211 static int open_mem(struct inode * inode, struct file * filp)
 212 {
 213         return capable(CAP_SYS_RAWIO) ? 0 : -EPERM;
 214 }
 215
 216 const struct file_operations mem_fops = {
 217         .llseek         = memory_lseek,
 218         .read           = read_mem,
 219         .write          = write_mem,
 220         .mmap           = xen_mmap_mem,
 221         .open           = open_mem,
 222 };