drivers/xen/privcmd/privcmd.c

   1 /******************************************************************************
   2  * privcmd.c
   3  *
   4  * Interface to privileged domain-0 commands.
   5  *
   6  * Copyright (c) 2002-2004, K A Fraser, B Dragovic
   7  */
   8
   9 #include <linux/kernel.h>
  10 #include <linux/sched.h>
  11 #include <linux/slab.h>
  12 #include <linux/string.h>
  13 #include <linux/errno.h>
  14 #include <linux/mm.h>
  15 #include <linux/mman.h>
  16 #include <linux/swap.h>
  17 #include <linux/smp_lock.h>
  18 #include <linux/highmem.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/seq_file.h>
  21 #include <linux/kthread.h>
  22 #include <asm/hypervisor.h>
  23
  24 #include <asm/pgalloc.h>
  25 #include <asm/pgtable.h>
  26 #include <asm/uaccess.h>
  27 #include <asm/tlb.h>
  28 #include <asm/hypervisor.h>
  29 #include <xen/public/privcmd.h>
  30 #include <xen/interface/xen.h>
  31 #include <xen/interface/dom0_ops.h>
  32 #include <xen/xen_proc.h>
  33
  34 static struct proc_dir_entry *privcmd_intf;
  35 static struct proc_dir_entry *capabilities_intf;
  36
  37 static int privcmd_ioctl(struct inode *inode, struct file *file,
  38                          unsigned int cmd, unsigned long data)
  39 {
  40         int ret = -ENOSYS;
  41         void __user *udata = (void __user *) data;
  42
  43         switch (cmd) {
  44         case IOCTL_PRIVCMD_HYPERCALL: {
  45                 privcmd_hypercall_t hypercall;
  46
  47                 if (copy_from_user(&hypercall, udata, sizeof(hypercall)))
  48                         return -EFAULT;
  49
  50 #if defined(__i386__)
  51                 __asm__ __volatile__ (
  52                         "pushl %%ebx; pushl %%ecx; pushl %%edx; "
  53                         "pushl %%esi; pushl %%edi; "
  54                         "movl  8(%%eax),%%ebx ;"
  55                         "movl 16(%%eax),%%ecx ;"
  56                         "movl 24(%%eax),%%edx ;"
  57                         "movl 32(%%eax),%%esi ;"
  58                         "movl 40(%%eax),%%edi ;"
  59                         "movl   (%%eax),%%eax ;"
  60                         "shll $5,%%eax ;"
  61                         "addl $hypercall_page,%%eax ;"
  62                         "call *%%eax ;"
  63                         "popl %%edi; popl %%esi; popl %%edx; "
  64                         "popl %%ecx; popl %%ebx"
  65                         : "=a" (ret) : "0" (&hypercall) : "memory" );
  66 #elif defined (__x86_64__)
  67                 {
  68                         long ign1, ign2, ign3;
  69                         __asm__ __volatile__ (
  70                                 "movq %8,%%r10; movq %9,%%r8;"
  71                                 "shlq $5,%%rax ;"
  72                                 "addq $hypercall_page,%%rax ;"
  73                                 "call *%%rax"
  74                                 : "=a" (ret), "=D" (ign1),
  75                                   "=S" (ign2), "=d" (ign3)
  76                                 : "0" ((unsigned long)hypercall.op),
  77                                 "1" ((unsigned long)hypercall.arg[0]),
  78                                 "2" ((unsigned long)hypercall.arg[1]),
  79                                 "3" ((unsigned long)hypercall.arg[2]),
  80                                 "g" ((unsigned long)hypercall.arg[3]),
  81                                 "g" ((unsigned long)hypercall.arg[4])
  82                                 : "r8", "r10", "memory" );
  83                 }
  84 #elif defined (__ia64__)
  85                 __asm__ __volatile__ (
  86                         ";; mov r14=%2; mov r15=%3; "
  87                         "mov r16=%4; mov r17=%5; mov r18=%6;"
  88                         "mov r2=%1; break 0x1000;; mov %0=r8 ;;"
  89                         : "=r" (ret)
  90                         : "r" (hypercall.op),
  91                         "r" (hypercall.arg[0]),
  92                         "r" (hypercall.arg[1]),
  93                         "r" (hypercall.arg[2]),
  94                         "r" (hypercall.arg[3]),
  95                         "r" (hypercall.arg[4])
  96                         : "r14","r15","r16","r17","r18","r2","r8","memory");
  97 #endif
  98         }
  99         break;
 100
 101         case IOCTL_PRIVCMD_MMAP: {
 102                 privcmd_mmap_t mmapcmd;
 103                 privcmd_mmap_entry_t msg;
 104                 privcmd_mmap_entry_t __user *p;
 105                 struct mm_struct *mm = current->mm;
 106                 struct vm_area_struct *vma;
 107                 unsigned long va;
 108                 int i, rc;
 109
 110                 if (!is_initial_xendomain())
 111                         return -EPERM;
 112
 113                 if (copy_from_user(&mmapcmd, udata, sizeof(mmapcmd)))
 114                         return -EFAULT;
 115
 116                 p = mmapcmd.entry;
 117                 if (copy_from_user(&msg, p, sizeof(msg)))
 118                         return -EFAULT;
 119
 120                 down_read(&mm->mmap_sem);
 121
 122                 vma = find_vma(mm, msg.va);
 123                 rc = -EINVAL;
 124                 if (!vma || (msg.va != vma->vm_start) || vma->vm_private_data)
 125                         goto mmap_out;
 126
 127                 /* Mapping is a one-shot operation per vma. */
 128                 vma->vm_private_data = (void *)1;
 129
 130                 va = vma->vm_start;
 131
 132                 for (i = 0; i < mmapcmd.num; i++) {
 133                         rc = -EFAULT;
 134                         if (copy_from_user(&msg, p, sizeof(msg)))
 135                                 goto mmap_out;
 136
 137                         /* Do not allow range to wrap the address space. */
 138                         rc = -EINVAL;
 139                         if ((msg.npages > (INT_MAX >> PAGE_SHIFT)) ||
 140                             ((unsigned long)(msg.npages << PAGE_SHIFT) >= -va))
 141                                 goto mmap_out;
 142
 143                         /* Range chunks must be contiguous in va space. */
 144                         if ((msg.va != va) ||
 145                             ((msg.va+(msg.npages<<PAGE_SHIFT)) > vma->vm_end))
 146                                 goto mmap_out;
 147
 148                         if ((rc = direct_remap_pfn_range(
 149                                 vma,
 150                                 msg.va & PAGE_MASK,
 151                                 msg.mfn,
 152                                 msg.npages << PAGE_SHIFT,
 153                                 vma->vm_page_prot,
 154                                 mmapcmd.dom)) < 0)
 155                                 goto mmap_out;
 156
 157                         p++;
 158                         va += msg.npages << PAGE_SHIFT;
 159                 }
 160
 161                 rc = 0;
 162
 163         mmap_out:
 164                 up_read(&mm->mmap_sem);
 165                 ret = rc;
 166         }
 167         break;
 168
 169         case IOCTL_PRIVCMD_MMAPBATCH: {
 170                 privcmd_mmapbatch_t m;
 171                 struct mm_struct *mm = current->mm;
 172                 struct vm_area_struct *vma;
 173                 xen_pfn_t __user *p;
 174                 unsigned long addr, mfn;
 175                 int i;
 176
 177                 if (!is_initial_xendomain())
 178                         return -EPERM;
 179
 180                 if (copy_from_user(&m, udata, sizeof(m)))
 181                         return -EFAULT;
 182
 183                 if ((m.num <= 0) || (m.num > (INT_MAX >> PAGE_SHIFT)))
 184                         return -EINVAL;
 185
 186                 down_read(&mm->mmap_sem);
 187
 188                 vma = find_vma(mm, m.addr);
 189                 if (!vma ||
 190                     (m.addr != vma->vm_start) ||
 191                     ((m.addr + (m.num<<PAGE_SHIFT)) != vma->vm_end) ||
 192                     vma->vm_private_data) {
 193                         up_read(&mm->mmap_sem);
 194                         return -EINVAL;
 195                 }
 196
 197                 /* Mapping is a one-shot operation per vma. */
 198                 vma->vm_private_data = (void *)1;
 199
 200                 p = m.arr;
 201                 addr = m.addr;
 202                 for (i = 0; i < m.num; i++, addr += PAGE_SIZE, p++) {
 203                         if (get_user(mfn, p)) {
 204                                 up_read(&mm->mmap_sem);
 205                                 return -EFAULT;
 206                         }
 207
 208                         ret = direct_remap_pfn_range(vma, addr & PAGE_MASK,
 209                                                      mfn, PAGE_SIZE,
 210                                                      vma->vm_page_prot, m.dom);
 211                         if (ret < 0)
 212                                 put_user(0xF0000000 | mfn, p);
 213                 }
 214
 215                 up_read(&mm->mmap_sem);
 216                 ret = 0;
 217         }
 218         break;
 219
 220         default:
 221                 ret = -EINVAL;
 222                 break;
 223         }
 224
 225         return ret;
 226 }
 227
 228 #ifndef HAVE_ARCH_PRIVCMD_MMAP
 229 static struct page *privcmd_nopage(struct vm_area_struct *vma,
 230                                    unsigned long address,
 231                                    int *type)
 232 {
 233         return NOPAGE_SIGBUS;
 234 }
 235
 236 static struct vm_operations_struct privcmd_vm_ops = {
 237         .nopage = privcmd_nopage
 238 };
 239
 240 static int privcmd_mmap(struct file * file, struct vm_area_struct * vma)
 241 {
 242         /* Unsupported for auto-translate guests. */
 243         if (xen_feature(XENFEAT_auto_translated_physmap))
 244                 return -ENOSYS;
 245
 246         /* DONTCOPY is essential for Xen as copy_page_range is broken. */
 247         vma->vm_flags |= VM_RESERVED | VM_IO | VM_DONTCOPY | VM_PFNMAP;
 248         vma->vm_ops = &privcmd_vm_ops;
 249         vma->vm_private_data = NULL;
 250
 251         return 0;
 252 }
 253 #endif
 254
 255 static struct file_operations privcmd_file_ops = {
 256         .ioctl = privcmd_ioctl,
 257         .mmap  = privcmd_mmap,
 258 };
 259
 260 static int capabilities_read(char *page, char **start, off_t off,
 261                              int count, int *eof, void *data)
 262 {
 263         int len = 0;
 264         *page = 0;
 265
 266         if (is_initial_xendomain())
 267                 len = sprintf( page, "control_d\n" );
 268
 269         *eof = 1;
 270         return len;
 271 }
 272
 273 static int __init privcmd_init(void)
 274 {
 275         if (!is_running_on_xen())
 276                 return -ENODEV;
 277
 278         privcmd_intf = create_xen_proc_entry("privcmd", 0400);
 279         if (privcmd_intf != NULL)
 280                 privcmd_intf->proc_fops = &privcmd_file_ops;
 281
 282         capabilities_intf = create_xen_proc_entry("capabilities", 0400 );
 283         if (capabilities_intf != NULL)
 284                 capabilities_intf->read_proc = capabilities_read;
 285
 286         return 0;
 287 }
 288
 289 __initcall(privcmd_init);