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 #ifndef HAVE_ARCH_PRIVCMD_MMAP
  38 static int privcmd_enforce_singleshot_mapping(struct vm_area_struct *vma);
  39 #endif
  40
  41 static int privcmd_ioctl(struct inode *inode, struct file *file,
  42                          unsigned int cmd, unsigned long data)
  43 {
  44         int ret = -ENOSYS;
  45         void __user *udata = (void __user *) data;
  46
  47         switch (cmd) {
  48         case IOCTL_PRIVCMD_HYPERCALL: {
  49                 privcmd_hypercall_t hypercall;
  50
  51                 if (copy_from_user(&hypercall, udata, sizeof(hypercall)))
  52                         return -EFAULT;
  53
  54 #if defined(__i386__)
  55                 __asm__ __volatile__ (
  56                         "pushl %%ebx; pushl %%ecx; pushl %%edx; "
  57                         "pushl %%esi; pushl %%edi; "
  58                         "movl  8(%%eax),%%ebx ;"
  59                         "movl 16(%%eax),%%ecx ;"
  60                         "movl 24(%%eax),%%edx ;"
  61                         "movl 32(%%eax),%%esi ;"
  62                         "movl 40(%%eax),%%edi ;"
  63                         "movl   (%%eax),%%eax ;"
  64                         "shll $5,%%eax ;"
  65                         "addl $hypercall_page,%%eax ;"
  66                         "call *%%eax ;"
  67                         "popl %%edi; popl %%esi; popl %%edx; "
  68                         "popl %%ecx; popl %%ebx"
  69                         : "=a" (ret) : "0" (&hypercall) : "memory" );
  70 #elif defined (__x86_64__)
  71                 {
  72                         long ign1, ign2, ign3;
  73                         __asm__ __volatile__ (
  74                                 "movq %8,%%r10; movq %9,%%r8;"
  75                                 "shlq $5,%%rax ;"
  76                                 "addq $hypercall_page,%%rax ;"
  77                                 "call *%%rax"
  78                                 : "=a" (ret), "=D" (ign1),
  79                                   "=S" (ign2), "=d" (ign3)
  80                                 : "0" ((unsigned long)hypercall.op),
  81                                 "1" ((unsigned long)hypercall.arg[0]),
  82                                 "2" ((unsigned long)hypercall.arg[1]),
  83                                 "3" ((unsigned long)hypercall.arg[2]),
  84                                 "g" ((unsigned long)hypercall.arg[3]),
  85                                 "g" ((unsigned long)hypercall.arg[4])
  86                                 : "r8", "r10", "memory" );
  87                 }
  88 #elif defined (__ia64__)
  89                 __asm__ __volatile__ (
  90                         ";; mov r14=%2; mov r15=%3; "
  91                         "mov r16=%4; mov r17=%5; mov r18=%6;"
  92                         "mov r2=%1; break 0x1000;; mov %0=r8 ;;"
  93                         : "=r" (ret)
  94                         : "r" (hypercall.op),
  95                         "r" (hypercall.arg[0]),
  96                         "r" (hypercall.arg[1]),
  97                         "r" (hypercall.arg[2]),
  98                         "r" (hypercall.arg[3]),
  99                         "r" (hypercall.arg[4])
 100                         : "r14","r15","r16","r17","r18","r2","r8","memory");
 101 #endif
 102         }
 103         break;
 104
 105         case IOCTL_PRIVCMD_MMAP: {
 106                 privcmd_mmap_t mmapcmd;
 107                 privcmd_mmap_entry_t msg;
 108                 privcmd_mmap_entry_t __user *p;
 109                 struct mm_struct *mm = current->mm;
 110                 struct vm_area_struct *vma;
 111                 unsigned long va;
 112                 int i, rc;
 113
 114                 if (!is_initial_xendomain())
 115                         return -EPERM;
 116
 117                 if (copy_from_user(&mmapcmd, udata, sizeof(mmapcmd)))
 118                         return -EFAULT;
 119
 120                 p = mmapcmd.entry;
 121                 if (copy_from_user(&msg, p, sizeof(msg)))
 122                         return -EFAULT;
 123
 124                 down_read(&mm->mmap_sem);
 125
 126                 vma = find_vma(mm, msg.va);
 127                 rc = -EINVAL;
 128                 if (!vma || (msg.va != vma->vm_start) ||
 129                     !privcmd_enforce_singleshot_mapping(vma))
 130                         goto mmap_out;
 131
 132                 va = vma->vm_start;
 133
 134                 for (i = 0; i < mmapcmd.num; i++) {
 135                         rc = -EFAULT;
 136                         if (copy_from_user(&msg, p, sizeof(msg)))
 137                                 goto mmap_out;
 138
 139                         /* Do not allow range to wrap the address space. */
 140                         rc = -EINVAL;
 141                         if ((msg.npages > (LONG_MAX >> PAGE_SHIFT)) ||
 142                             ((unsigned long)(msg.npages << PAGE_SHIFT) >= -va))
 143                                 goto mmap_out;
 144
 145                         /* Range chunks must be contiguous in va space. */
 146                         if ((msg.va != va) ||
 147                             ((msg.va+(msg.npages<<PAGE_SHIFT)) > vma->vm_end))
 148                                 goto mmap_out;
 149
 150                         if ((rc = direct_remap_pfn_range(
 151                                 vma,
 152                                 msg.va & PAGE_MASK,
 153                                 msg.mfn,
 154                                 msg.npages << PAGE_SHIFT,
 155                                 vma->vm_page_prot,
 156                                 mmapcmd.dom)) < 0)
 157                                 goto mmap_out;
 158
 159                         p++;
 160                         va += msg.npages << PAGE_SHIFT;
 161                 }
 162
 163                 rc = 0;
 164
 165         mmap_out:
 166                 up_read(&mm->mmap_sem);
 167                 ret = rc;
 168         }
 169         break;
 170
 171         case IOCTL_PRIVCMD_MMAPBATCH: {
 172                 privcmd_mmapbatch_t m;
 173                 struct mm_struct *mm = current->mm;
 174                 struct vm_area_struct *vma;
 175                 xen_pfn_t __user *p;
 176                 unsigned long addr, mfn;
 177                 int i;
 178
 179                 if (!is_initial_xendomain())
 180                         return -EPERM;
 181
 182                 if (copy_from_user(&m, udata, sizeof(m)))
 183                         return -EFAULT;
 184
 185                 if ((m.num <= 0) || (m.num > (LONG_MAX >> PAGE_SHIFT)))
 186                         return -EINVAL;
 187
 188                 down_read(&mm->mmap_sem);
 189
 190                 vma = find_vma(mm, m.addr);
 191                 if (!vma ||
 192                     (m.addr != vma->vm_start) ||
 193                     ((m.addr + ((unsigned long)m.num<<PAGE_SHIFT)) !=
 194                      vma->vm_end) ||
 195                     !privcmd_enforce_singleshot_mapping(vma)) {
 196                         up_read(&mm->mmap_sem);
 197                         return -EINVAL;
 198                 }
 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
 254 static int privcmd_enforce_singleshot_mapping(struct vm_area_struct *vma)
 255 {
 256         return (xchg(&vma->vm_private_data, (void *)1) == NULL);
 257 }
 258 #endif
 259
 260 static struct file_operations privcmd_file_ops = {
 261         .ioctl = privcmd_ioctl,
 262         .mmap  = privcmd_mmap,
 263 };
 264
 265 static int capabilities_read(char *page, char **start, off_t off,
 266                              int count, int *eof, void *data)
 267 {
 268         int len = 0;
 269         *page = 0;
 270
 271         if (is_initial_xendomain())
 272                 len = sprintf( page, "control_d\n" );
 273
 274         *eof = 1;
 275         return len;
 276 }
 277
 278 static int __init privcmd_init(void)
 279 {
 280         if (!is_running_on_xen())
 281                 return -ENODEV;
 282
 283         privcmd_intf = create_xen_proc_entry("privcmd", 0400);
 284         if (privcmd_intf != NULL)
 285                 privcmd_intf->proc_fops = &privcmd_file_ops;
 286
 287         capabilities_intf = create_xen_proc_entry("capabilities", 0400 );
 288         if (capabilities_intf != NULL)
 289                 capabilities_intf->read_proc = capabilities_read;
 290
 291         return 0;
 292 }
 293
 294 __initcall(privcmd_init);