drivers/lguest/core.c

   1 /*P:400 This contains run_guest() which actually calls into the Host<->Guest
   2  * Switcher and analyzes the return, such as determining if the Guest wants the
   3  * Host to do something.  This file also contains useful helper routines, and a
   4  * couple of non-obvious setup and teardown pieces which were implemented after
   5  * days of debugging pain. :*/
   6 #include <linux/module.h>
   7 #include <linux/stringify.h>
   8 #include <linux/stddef.h>
   9 #include <linux/io.h>
  10 #include <linux/mm.h>
  11 #include <linux/vmalloc.h>
  12 #include <linux/cpu.h>
  13 #include <linux/freezer.h>
  14 #include <linux/highmem.h>
  15 #include <asm/paravirt.h>
  16 #include <asm/pgtable.h>
  17 #include <asm/uaccess.h>
  18 #include <asm/poll.h>
  19 #include <asm/asm-offsets.h>
  20 #include "lg.h"
  21
  22
  23 static struct vm_struct *switcher_vma;
  24 static struct page **switcher_page;
  25
  26 /* This One Big lock protects all inter-guest data structures. */
  27 DEFINE_MUTEX(lguest_lock);
  28
  29 /*H:010 We need to set up the Switcher at a high virtual address.  Remember the
  30  * Switcher is a few hundred bytes of assembler code which actually changes the
  31  * CPU to run the Guest, and then changes back to the Host when a trap or
  32  * interrupt happens.
  33  *
  34  * The Switcher code must be at the same virtual address in the Guest as the
  35  * Host since it will be running as the switchover occurs.
  36  *
  37  * Trying to map memory at a particular address is an unusual thing to do, so
  38  * it's not a simple one-liner. */
  39 static __init int map_switcher(void)
  40 {
  41         int i, err;
  42         struct page **pagep;
  43
  44         /*
  45          * Map the Switcher in to high memory.
  46          *
  47          * It turns out that if we choose the address 0xFFC00000 (4MB under the
  48          * top virtual address), it makes setting up the page tables really
  49          * easy.
  50          */
  51
  52         /* We allocate an array of "struct page"s.  map_vm_area() wants the
  53          * pages in this form, rather than just an array of pointers. */
  54         switcher_page = kmalloc(sizeof(switcher_page[0])*TOTAL_SWITCHER_PAGES,
  55                                 GFP_KERNEL);
  56         if (!switcher_page) {
  57                 err = -ENOMEM;
  58                 goto out;
  59         }
  60
  61         /* Now we actually allocate the pages.  The Guest will see these pages,
  62          * so we make sure they're zeroed. */
  63         for (i = 0; i < TOTAL_SWITCHER_PAGES; i++) {
  64                 unsigned long addr = get_zeroed_page(GFP_KERNEL);
  65                 if (!addr) {
  66                         err = -ENOMEM;
  67                         goto free_some_pages;
  68                 }
  69                 switcher_page[i] = virt_to_page(addr);
  70         }
  71
  72         /* Now we reserve the "virtual memory area" we want: 0xFFC00000
  73          * (SWITCHER_ADDR).  We might not get it in theory, but in practice
  74          * it's worked so far. */
  75         switcher_vma = __get_vm_area(TOTAL_SWITCHER_PAGES * PAGE_SIZE,
  76                                        VM_ALLOC, SWITCHER_ADDR, VMALLOC_END);
  77         if (!switcher_vma) {
  78                 err = -ENOMEM;
  79                 printk("lguest: could not map switcher pages high\n");
  80                 goto free_pages;
  81         }
  82
  83         /* This code actually sets up the pages we've allocated to appear at
  84          * SWITCHER_ADDR.  map_vm_area() takes the vma we allocated above, the
  85          * kind of pages we're mapping (kernel pages), and a pointer to our
  86          * array of struct pages.  It increments that pointer, but we don't
  87          * care. */
  88         pagep = switcher_page;
  89         err = map_vm_area(switcher_vma, PAGE_KERNEL, &pagep);
  90         if (err) {
  91                 printk("lguest: map_vm_area failed: %i\n", err);
  92                 goto free_vma;
  93         }
  94
  95         /* Now the Switcher is mapped at the right address, we can't fail!
  96          * Copy in the compiled-in Switcher code (from <arch>_switcher.S). */
  97         memcpy(switcher_vma->addr, start_switcher_text,
  98                end_switcher_text - start_switcher_text);
  99
 100         printk(KERN_INFO "lguest: mapped switcher at %p\n",
 101                switcher_vma->addr);
 102         /* And we succeeded... */
 103         return 0;
 104
 105 free_vma:
 106         vunmap(switcher_vma->addr);
 107 free_pages:
 108         i = TOTAL_SWITCHER_PAGES;
 109 free_some_pages:
 110         for (--i; i >= 0; i--)
 111                 __free_pages(switcher_page[i], 0);
 112         kfree(switcher_page);
 113 out:
 114         return err;
 115 }
 116 /*:*/
 117
 118 /* Cleaning up the mapping when the module is unloaded is almost...
 119  * too easy. */
 120 static void unmap_switcher(void)
 121 {
 122         unsigned int i;
 123
 124         /* vunmap() undoes *both* map_vm_area() and __get_vm_area(). */
 125         vunmap(switcher_vma->addr);
 126         /* Now we just need to free the pages we copied the switcher into */
 127         for (i = 0; i < TOTAL_SWITCHER_PAGES; i++)
 128                 __free_pages(switcher_page[i], 0);
 129 }
 130
 131 /*L:305
 132  * Dealing With Guest Memory.
 133  *
 134  * When the Guest gives us (what it thinks is) a physical address, we can use
 135  * the normal copy_from_user() & copy_to_user() on the corresponding place in
 136  * the memory region allocated by the Launcher.
 137  *
 138  * But we can't trust the Guest: it might be trying to access the Launcher
 139  * code.  We have to check that the range is below the pfn_limit the Launcher
 140  * gave us.  We have to make sure that addr + len doesn't give us a false
 141  * positive by overflowing, too. */
 142 int lguest_address_ok(const struct lguest *lg,
 143                       unsigned long addr, unsigned long len)
 144 {
 145         return (addr+len) / PAGE_SIZE < lg->pfn_limit && (addr+len >= addr);
 146 }
 147
 148 /* This is a convenient routine to get a 32-bit value from the Guest (a very
 149  * common operation).  Here we can see how useful the kill_lguest() routine we
 150  * met in the Launcher can be: we return a random value (0) instead of needing
 151  * to return an error. */
 152 u32 lgread_u32(struct lguest *lg, unsigned long addr)
 153 {
 154         u32 val = 0;
 155
 156         /* Don't let them access lguest binary. */
 157         if (!lguest_address_ok(lg, addr, sizeof(val))
 158             || get_user(val, (u32 *)(lg->mem_base + addr)) != 0)
 159                 kill_guest(lg, "bad read address %#lx: pfn_limit=%u membase=%p", addr, lg->pfn_limit, lg->mem_base);
 160         return val;
 161 }
 162
 163 /* Same thing for writing a value. */
 164 void lgwrite_u32(struct lguest *lg, unsigned long addr, u32 val)
 165 {
 166         if (!lguest_address_ok(lg, addr, sizeof(val))
 167             || put_user(val, (u32 *)(lg->mem_base + addr)) != 0)
 168                 kill_guest(lg, "bad write address %#lx", addr);
 169 }
 170
 171 /* This routine is more generic, and copies a range of Guest bytes into a
 172  * buffer.  If the copy_from_user() fails, we fill the buffer with zeroes, so
 173  * the caller doesn't end up using uninitialized kernel memory. */
 174 void lgread(struct lguest *lg, void *b, unsigned long addr, unsigned bytes)
 175 {
 176         if (!lguest_address_ok(lg, addr, bytes)
 177             || copy_from_user(b, lg->mem_base + addr, bytes) != 0) {
 178                 /* copy_from_user should do this, but as we rely on it... */
 179                 memset(b, 0, bytes);
 180                 kill_guest(lg, "bad read address %#lx len %u", addr, bytes);
 181         }
 182 }
 183
 184 /* Similarly, our generic routine to copy into a range of Guest bytes. */
 185 void lgwrite(struct lguest *lg, unsigned long addr, const void *b,
 186              unsigned bytes)
 187 {
 188         if (!lguest_address_ok(lg, addr, bytes)
 189             || copy_to_user(lg->mem_base + addr, b, bytes) != 0)
 190                 kill_guest(lg, "bad write address %#lx len %u", addr, bytes);
 191 }
 192 /* (end of memory access helper routines) :*/
 193
 194 /*H:030 Let's jump straight to the the main loop which runs the Guest.
 195  * Remember, this is called by the Launcher reading /dev/lguest, and we keep
 196  * going around and around until something interesting happens. */
 197 int run_guest(struct lguest *lg, unsigned long __user *user)
 198 {
 199         /* We stop running once the Guest is dead. */
 200         while (!lg->dead) {
 201                 /* First we run any hypercalls the Guest wants done: either in
 202                  * the hypercall ring in "struct lguest_data", or directly by
 203                  * using int 31 (LGUEST_TRAP_ENTRY). */
 204                 do_hypercalls(lg);
 205                 /* It's possible the Guest did a SEND_DMA hypercall to the
 206                  * Launcher, in which case we return from the read() now. */
 207                 if (lg->dma_is_pending) {
 208                         if (put_user(lg->pending_dma, user) ||
 209                             put_user(lg->pending_key, user+1))
 210                                 return -EFAULT;
 211                         return sizeof(unsigned long)*2;
 212                 }
 213
 214                 /* Check for signals */
 215                 if (signal_pending(current))
 216                         return -ERESTARTSYS;
 217
 218                 /* If Waker set break_out, return to Launcher. */
 219                 if (lg->break_out)
 220                         return -EAGAIN;
 221
 222                 /* Check if there are any interrupts which can be delivered
 223                  * now: if so, this sets up the hander to be executed when we
 224                  * next run the Guest. */
 225                 maybe_do_interrupt(lg);
 226
 227                 /* All long-lived kernel loops need to check with this horrible
 228                  * thing called the freezer.  If the Host is trying to suspend,
 229                  * it stops us. */
 230                 try_to_freeze();
 231
 232                 /* Just make absolutely sure the Guest is still alive.  One of
 233                  * those hypercalls could have been fatal, for example. */
 234                 if (lg->dead)
 235                         break;
 236
 237                 /* If the Guest asked to be stopped, we sleep.  The Guest's
 238                  * clock timer or LHCALL_BREAK from the Waker will wake us. */
 239                 if (lg->halted) {
 240                         set_current_state(TASK_INTERRUPTIBLE);
 241                         schedule();
 242                         continue;
 243                 }
 244
 245                 /* OK, now we're ready to jump into the Guest.  First we put up
 246                  * the "Do Not Disturb" sign: */
 247                 local_irq_disable();
 248
 249                 /* Actually run the Guest until something happens. */
 250                 lguest_arch_run_guest(lg);
 251
 252                 /* Now we're ready to be interrupted or moved to other CPUs */
 253                 local_irq_enable();
 254
 255                 /* Now we deal with whatever happened to the Guest. */
 256                 lguest_arch_handle_trap(lg);
 257         }
 258
 259         /* The Guest is dead => "No such file or directory" */
 260         return -ENOENT;
 261 }
 262
 263 /*H:000
 264  * Welcome to the Host!
 265  *
 266  * By this point your brain has been tickled by the Guest code and numbed by
 267  * the Launcher code; prepare for it to be stretched by the Host code.  This is
 268  * the heart.  Let's begin at the initialization routine for the Host's lg
 269  * module.
 270  */
 271 static int __init init(void)
 272 {
 273         int err;
 274
 275         /* Lguest can't run under Xen, VMI or itself.  It does Tricky Stuff. */
 276         if (paravirt_enabled()) {
 277                 printk("lguest is afraid of %s\n", pv_info.name);
 278                 return -EPERM;
 279         }
 280
 281         /* First we put the Switcher up in very high virtual memory. */
 282         err = map_switcher();
 283         if (err)
 284                 return err;
 285
 286         /* Now we set up the pagetable implementation for the Guests. */
 287         err = init_pagetables(switcher_page, SHARED_SWITCHER_PAGES);
 288         if (err) {
 289                 unmap_switcher();
 290                 return err;
 291         }
 292
 293         /* The I/O subsystem needs some things initialized. */
 294         lguest_io_init();
 295
 296         /* /dev/lguest needs to be registered. */
 297         err = lguest_device_init();
 298         if (err) {
 299                 free_pagetables();
 300                 unmap_switcher();
 301                 return err;
 302         }
 303
 304         /* Finally we do some architecture-specific setup. */
 305         lguest_arch_host_init();
 306
 307         /* All good! */
 308         return 0;
 309 }
 310
 311 /* Cleaning up is just the same code, backwards.  With a little French. */
 312 static void __exit fini(void)
 313 {
 314         lguest_device_remove();
 315         free_pagetables();
 316         unmap_switcher();
 317
 318         lguest_arch_host_fini();
 319 }
 320 /*:*/
 321
 322 /* The Host side of lguest can be a module.  This is a nice way for people to
 323  * play with it.  */
 324 module_init(init);
 325 module_exit(fini);
 326 MODULE_LICENSE("GPL");
 327 MODULE_AUTHOR("Rusty Russell <rusty@rustcorp.com.au>");