drivers/lguest/core.c

   1 /*P:400
   2  * This contains run_guest() which actually calls into the Host<->Guest
   3  * Switcher and analyzes the return, such as determining if the Guest wants the
   4  * Host to do something.  This file also contains useful helper routines.
   5 :*/
   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 <linux/slab.h>
  16 #include <asm/paravirt.h>
  17 #include <asm/pgtable.h>
  18 #include <linux/uaccess.h>
  19 #include <asm/poll.h>
  20 #include <asm/asm-offsets.h>
  21 #include "lg.h"
  22
  23 unsigned long switcher_addr;
  24 struct page **lg_switcher_pages;
  25 static struct vm_struct *switcher_text_vma;
  26 static struct vm_struct *switcher_stacks_vma;
  27
  28 /* This One Big lock protects all inter-guest data structures. */
  29 DEFINE_MUTEX(lguest_lock);
  30
  31 /*H:010
  32  * We need to set up the Switcher at a high virtual address.  Remember the
  33  * Switcher is a few hundred bytes of assembler code which actually changes the
  34  * CPU to run the Guest, and then changes back to the Host when a trap or
  35  * interrupt happens.
  36  *
  37  * The Switcher code must be at the same virtual address in the Guest as the
  38  * Host since it will be running as the switchover occurs.
  39  *
  40  * Trying to map memory at a particular address is an unusual thing to do, so
  41  * it's not a simple one-liner.
  42  */
  43 static __init int map_switcher(void)
  44 {
  45         int i, err;
  46
  47         /*
  48          * Map the Switcher in to high memory.
  49          *
  50          * It turns out that if we choose the address 0xFFC00000 (4MB under the
  51          * top virtual address), it makes setting up the page tables really
  52          * easy.
  53          */
  54
  55         /* We assume Switcher text fits into a single page. */
  56         if (end_switcher_text - start_switcher_text > PAGE_SIZE) {
  57                 printk(KERN_ERR "lguest: switcher text too large (%zu)\n",
  58                        end_switcher_text - start_switcher_text);
  59                 return -EINVAL;
  60         }
  61
  62         /*
  63          * We allocate an array of struct page pointers.  map_vm_area() wants
  64          * this, rather than just an array of pages.
  65          */
  66         lg_switcher_pages = kmalloc(sizeof(lg_switcher_pages[0])
  67                                     * TOTAL_SWITCHER_PAGES,
  68                                     GFP_KERNEL);
  69         if (!lg_switcher_pages) {
  70                 err = -ENOMEM;
  71                 goto out;
  72         }
  73
  74         /*
  75          * Now we actually allocate the pages.  The Guest will see these pages,
  76          * so we make sure they're zeroed.
  77          */
  78         for (i = 0; i < TOTAL_SWITCHER_PAGES; i++) {
  79                 lg_switcher_pages[i] = alloc_page(GFP_KERNEL|__GFP_ZERO);
  80                 if (!lg_switcher_pages[i]) {
  81                         err = -ENOMEM;
  82                         goto free_some_pages;
  83                 }
  84         }
  85
  86         /*
  87          * Copy in the compiled-in Switcher code (from x86/switcher_32.S).
  88          * It goes in the first page, which we map in momentarily.
  89          */
  90         memcpy(kmap(lg_switcher_pages[0]), start_switcher_text,
  91                end_switcher_text - start_switcher_text);
  92         kunmap(lg_switcher_pages[0]);
  93
  94         /*
  95          * We place the Switcher underneath the fixmap area, which is the
  96          * highest virtual address we can get.  This is important, since we
  97          * tell the Guest it can't access this memory, so we want its ceiling
  98          * as high as possible.
  99          */
 100         switcher_addr = FIXADDR_START - TOTAL_SWITCHER_PAGES*PAGE_SIZE;
 101
 102         /*
 103          * Now we reserve the "virtual memory area"s we want.  We might
 104          * not get them in theory, but in practice it's worked so far.
 105          *
 106          * We want the switcher text to be read-only and executable, and
 107          * the stacks to be read-write and non-executable.
 108          */
 109         switcher_text_vma = __get_vm_area(PAGE_SIZE, VM_ALLOC|VM_NO_GUARD,
 110                                           switcher_addr,
 111                                           switcher_addr + PAGE_SIZE);
 112
 113         if (!switcher_text_vma) {
 114                 err = -ENOMEM;
 115                 printk("lguest: could not map switcher pages high\n");
 116                 goto free_pages;
 117         }
 118
 119         switcher_stacks_vma = __get_vm_area(SWITCHER_STACK_PAGES * PAGE_SIZE,
 120                                             VM_ALLOC|VM_NO_GUARD,
 121                                             switcher_addr + PAGE_SIZE,
 122                                             switcher_addr + TOTAL_SWITCHER_PAGES * PAGE_SIZE);
 123         if (!switcher_stacks_vma) {
 124                 err = -ENOMEM;
 125                 printk("lguest: could not map switcher pages high\n");
 126                 goto free_text_vma;
 127         }
 128
 129         /*
 130          * This code actually sets up the pages we've allocated to appear at
 131          * switcher_addr.  map_vm_area() takes the vma we allocated above, the
 132          * kind of pages we're mapping (kernel text pages and kernel writable
 133          * pages respectively), and a pointer to our array of struct pages.
 134          */
 135         err = map_vm_area(switcher_text_vma, PAGE_KERNEL_RX, lg_switcher_pages);
 136         if (err) {
 137                 printk("lguest: text map_vm_area failed: %i\n", err);
 138                 goto free_vmas;
 139         }
 140
 141         err = map_vm_area(switcher_stacks_vma, PAGE_KERNEL,
 142                           lg_switcher_pages + SWITCHER_TEXT_PAGES);
 143         if (err) {
 144                 printk("lguest: stacks map_vm_area failed: %i\n", err);
 145                 goto free_vmas;
 146         }
 147
 148         /*
 149          * Now the Switcher is mapped at the right address, we can't fail!
 150          */
 151         printk(KERN_INFO "lguest: mapped switcher at %p\n",
 152                switcher_text_vma->addr);
 153         /* And we succeeded... */
 154         return 0;
 155
 156 free_vmas:
 157         /* Undoes map_vm_area and __get_vm_area */
 158         vunmap(switcher_stacks_vma->addr);
 159 free_text_vma:
 160         vunmap(switcher_text_vma->addr);
 161 free_pages:
 162         i = TOTAL_SWITCHER_PAGES;
 163 free_some_pages:
 164         for (--i; i >= 0; i--)
 165                 __free_pages(lg_switcher_pages[i], 0);
 166         kfree(lg_switcher_pages);
 167 out:
 168         return err;
 169 }
 170 /*:*/
 171
 172 /* Cleaning up the mapping when the module is unloaded is almost... too easy. */
 173 static void unmap_switcher(void)
 174 {
 175         unsigned int i;
 176
 177         /* vunmap() undoes *both* map_vm_area() and __get_vm_area(). */
 178         vunmap(switcher_text_vma->addr);
 179         vunmap(switcher_stacks_vma->addr);
 180         /* Now we just need to free the pages we copied the switcher into */
 181         for (i = 0; i < TOTAL_SWITCHER_PAGES; i++)
 182                 __free_pages(lg_switcher_pages[i], 0);
 183         kfree(lg_switcher_pages);
 184 }
 185
 186 /*H:032
 187  * Dealing With Guest Memory.
 188  *
 189  * Before we go too much further into the Host, we need to grok the routines
 190  * we use to deal with Guest memory.
 191  *
 192  * When the Guest gives us (what it thinks is) a physical address, we can use
 193  * the normal copy_from_user() & copy_to_user() on the corresponding place in
 194  * the memory region allocated by the Launcher.
 195  *
 196  * But we can't trust the Guest: it might be trying to access the Launcher
 197  * code.  We have to check that the range is below the pfn_limit the Launcher
 198  * gave us.  We have to make sure that addr + len doesn't give us a false
 199  * positive by overflowing, too.
 200  */
 201 bool lguest_address_ok(const struct lguest *lg,
 202                        unsigned long addr, unsigned long len)
 203 {
 204         return addr+len <= lg->pfn_limit * PAGE_SIZE && (addr+len >= addr);
 205 }
 206
 207 /*
 208  * This routine copies memory from the Guest.  Here we can see how useful the
 209  * kill_lguest() routine we met in the Launcher can be: we return a random
 210  * value (all zeroes) instead of needing to return an error.
 211  */
 212 void __lgread(struct lg_cpu *cpu, void *b, unsigned long addr, unsigned bytes)
 213 {
 214         if (!lguest_address_ok(cpu->lg, addr, bytes)
 215             || copy_from_user(b, cpu->lg->mem_base + addr, bytes) != 0) {
 216                 /* copy_from_user should do this, but as we rely on it... */
 217                 memset(b, 0, bytes);
 218                 kill_guest(cpu, "bad read address %#lx len %u", addr, bytes);
 219         }
 220 }
 221
 222 /* This is the write (copy into Guest) version. */
 223 void __lgwrite(struct lg_cpu *cpu, unsigned long addr, const void *b,
 224                unsigned bytes)
 225 {
 226         if (!lguest_address_ok(cpu->lg, addr, bytes)
 227             || copy_to_user(cpu->lg->mem_base + addr, b, bytes) != 0)
 228                 kill_guest(cpu, "bad write address %#lx len %u", addr, bytes);
 229 }
 230 /*:*/
 231
 232 /*H:030
 233  * Let's jump straight to the the main loop which runs the Guest.
 234  * Remember, this is called by the Launcher reading /dev/lguest, and we keep
 235  * going around and around until something interesting happens.
 236  */
 237 int run_guest(struct lg_cpu *cpu, unsigned long __user *user)
 238 {
 239         /* If the launcher asked for a register with LHREQ_GETREG */
 240         if (cpu->reg_read) {
 241                 if (put_user(*cpu->reg_read, user))
 242                         return -EFAULT;
 243                 cpu->reg_read = NULL;
 244                 return sizeof(*cpu->reg_read);
 245         }
 246
 247         /* We stop running once the Guest is dead. */
 248         while (!cpu->lg->dead) {
 249                 unsigned int irq;
 250                 bool more;
 251
 252                 /* First we run any hypercalls the Guest wants done. */
 253                 if (cpu->hcall)
 254                         do_hypercalls(cpu);
 255
 256                 /* Do we have to tell the Launcher about a trap? */
 257                 if (cpu->pending.trap) {
 258                         if (copy_to_user(user, &cpu->pending,
 259                                          sizeof(cpu->pending)))
 260                                 return -EFAULT;
 261                         return sizeof(cpu->pending);
 262                 }
 263
 264                 /*
 265                  * All long-lived kernel loops need to check with this horrible
 266                  * thing called the freezer.  If the Host is trying to suspend,
 267                  * it stops us.
 268                  */
 269                 try_to_freeze();
 270
 271                 /* Check for signals */
 272                 if (signal_pending(current))
 273                         return -ERESTARTSYS;
 274
 275                 /*
 276                  * Check if there are any interrupts which can be delivered now:
 277                  * if so, this sets up the hander to be executed when we next
 278                  * run the Guest.
 279                  */
 280                 irq = interrupt_pending(cpu, &more);
 281                 if (irq < LGUEST_IRQS)
 282                         try_deliver_interrupt(cpu, irq, more);
 283
 284                 /*
 285                  * Just make absolutely sure the Guest is still alive.  One of
 286                  * those hypercalls could have been fatal, for example.
 287                  */
 288                 if (cpu->lg->dead)
 289                         break;
 290
 291                 /*
 292                  * If the Guest asked to be stopped, we sleep.  The Guest's
 293                  * clock timer will wake us.
 294                  */
 295                 if (cpu->halted) {
 296                         set_current_state(TASK_INTERRUPTIBLE);
 297                         /*
 298                          * Just before we sleep, make sure no interrupt snuck in
 299                          * which we should be doing.
 300                          */
 301                         if (interrupt_pending(cpu, &more) < LGUEST_IRQS)
 302                                 set_current_state(TASK_RUNNING);
 303                         else
 304                                 schedule();
 305                         continue;
 306                 }
 307
 308                 /*
 309                  * OK, now we're ready to jump into the Guest.  First we put up
 310                  * the "Do Not Disturb" sign:
 311                  */
 312                 local_irq_disable();
 313
 314                 /* Actually run the Guest until something happens. */
 315                 lguest_arch_run_guest(cpu);
 316
 317                 /* Now we're ready to be interrupted or moved to other CPUs */
 318                 local_irq_enable();
 319
 320                 /* Now we deal with whatever happened to the Guest. */
 321                 lguest_arch_handle_trap(cpu);
 322         }
 323
 324         /* Special case: Guest is 'dead' but wants a reboot. */
 325         if (cpu->lg->dead == ERR_PTR(-ERESTART))
 326                 return -ERESTART;
 327
 328         /* The Guest is dead => "No such file or directory" */
 329         return -ENOENT;
 330 }
 331
 332 /*H:000
 333  * Welcome to the Host!
 334  *
 335  * By this point your brain has been tickled by the Guest code and numbed by
 336  * the Launcher code; prepare for it to be stretched by the Host code.  This is
 337  * the heart.  Let's begin at the initialization routine for the Host's lg
 338  * module.
 339  */
 340 static int __init init(void)
 341 {
 342         int err;
 343
 344         /* Lguest can't run under Xen, VMI or itself.  It does Tricky Stuff. */
 345         if (get_kernel_rpl() != 0) {
 346                 printk("lguest is afraid of being a guest\n");
 347                 return -EPERM;
 348         }
 349
 350         /* First we put the Switcher up in very high virtual memory. */
 351         err = map_switcher();
 352         if (err)
 353                 goto out;
 354
 355         /* We might need to reserve an interrupt vector. */
 356         err = init_interrupts();
 357         if (err)
 358                 goto unmap;
 359
 360         /* /dev/lguest needs to be registered. */
 361         err = lguest_device_init();
 362         if (err)
 363                 goto free_interrupts;
 364
 365         /* Finally we do some architecture-specific setup. */
 366         lguest_arch_host_init();
 367
 368         /* All good! */
 369         return 0;
 370
 371 free_interrupts:
 372         free_interrupts();
 373 unmap:
 374         unmap_switcher();
 375 out:
 376         return err;
 377 }
 378
 379 /* Cleaning up is just the same code, backwards.  With a little French. */
 380 static void __exit fini(void)
 381 {
 382         lguest_device_remove();
 383         free_interrupts();
 384         unmap_switcher();
 385
 386         lguest_arch_host_fini();
 387 }
 388 /*:*/
 389
 390 /*
 391  * The Host side of lguest can be a module.  This is a nice way for people to
 392  * play with it.
 393  */
 394 module_init(init);
 395 module_exit(fini);
 396 MODULE_LICENSE("GPL");
 397 MODULE_AUTHOR("Rusty Russell <rusty@rustcorp.com.au>");