drivers/lguest/interrupts_and_traps.c

   1 /*P:800 Interrupts (traps) are complicated enough to earn their own file.
   2  * There are three classes of interrupts:
   3  *
   4  * 1) Real hardware interrupts which occur while we're running the Guest,
   5  * 2) Interrupts for virtual devices attached to the Guest, and
   6  * 3) Traps and faults from the Guest.
   7  *
   8  * Real hardware interrupts must be delivered to the Host, not the Guest.
   9  * Virtual interrupts must be delivered to the Guest, but we make them look
  10  * just like real hardware would deliver them.  Traps from the Guest can be set
  11  * up to go directly back into the Guest, but sometimes the Host wants to see
  12  * them first, so we also have a way of "reflecting" them into the Guest as if
  13  * they had been delivered to it directly. :*/
  14 #include <linux/uaccess.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/module.h>
  17 #include "lg.h"
  18
  19 /* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */
  20 static unsigned int syscall_vector = SYSCALL_VECTOR;
  21 module_param(syscall_vector, uint, 0444);
  22
  23 /* The address of the interrupt handler is split into two bits: */
  24 static unsigned long idt_address(u32 lo, u32 hi)
  25 {
  26         return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
  27 }
  28
  29 /* The "type" of the interrupt handler is a 4 bit field: we only support a
  30  * couple of types. */
  31 static int idt_type(u32 lo, u32 hi)
  32 {
  33         return (hi >> 8) & 0xF;
  34 }
  35
  36 /* An IDT entry can't be used unless the "present" bit is set. */
  37 static int idt_present(u32 lo, u32 hi)
  38 {
  39         return (hi & 0x8000);
  40 }
  41
  42 /* We need a helper to "push" a value onto the Guest's stack, since that's a
  43  * big part of what delivering an interrupt does. */
  44 static void push_guest_stack(struct lguest *lg, unsigned long *gstack, u32 val)
  45 {
  46         /* Stack grows upwards: move stack then write value. */
  47         *gstack -= 4;
  48         lgwrite(lg, *gstack, u32, val);
  49 }
  50
  51 /*H:210 The set_guest_interrupt() routine actually delivers the interrupt or
  52  * trap.  The mechanics of delivering traps and interrupts to the Guest are the
  53  * same, except some traps have an "error code" which gets pushed onto the
  54  * stack as well: the caller tells us if this is one.
  55  *
  56  * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
  57  * interrupt or trap.  It's split into two parts for traditional reasons: gcc
  58  * on i386 used to be frightened by 64 bit numbers.
  59  *
  60  * We set up the stack just like the CPU does for a real interrupt, so it's
  61  * identical for the Guest (and the standard "iret" instruction will undo
  62  * it). */
  63 static void set_guest_interrupt(struct lg_cpu *cpu, u32 lo, u32 hi, int has_err)
  64 {
  65         unsigned long gstack, origstack;
  66         u32 eflags, ss, irq_enable;
  67         unsigned long virtstack;
  68         struct lguest *lg = cpu->lg;
  69
  70         /* There are two cases for interrupts: one where the Guest is already
  71          * in the kernel, and a more complex one where the Guest is in
  72          * userspace.  We check the privilege level to find out. */
  73         if ((cpu->regs->ss&0x3) != GUEST_PL) {
  74                 /* The Guest told us their kernel stack with the SET_STACK
  75                  * hypercall: both the virtual address and the segment */
  76                 virtstack = cpu->esp1;
  77                 ss = cpu->ss1;
  78
  79                 origstack = gstack = guest_pa(lg, virtstack);
  80                 /* We push the old stack segment and pointer onto the new
  81                  * stack: when the Guest does an "iret" back from the interrupt
  82                  * handler the CPU will notice they're dropping privilege
  83                  * levels and expect these here. */
  84                 push_guest_stack(lg, &gstack, cpu->regs->ss);
  85                 push_guest_stack(lg, &gstack, cpu->regs->esp);
  86         } else {
  87                 /* We're staying on the same Guest (kernel) stack. */
  88                 virtstack = cpu->regs->esp;
  89                 ss = cpu->regs->ss;
  90
  91                 origstack = gstack = guest_pa(lg, virtstack);
  92         }
  93
  94         /* Remember that we never let the Guest actually disable interrupts, so
  95          * the "Interrupt Flag" bit is always set.  We copy that bit from the
  96          * Guest's "irq_enabled" field into the eflags word: we saw the Guest
  97          * copy it back in "lguest_iret". */
  98         eflags = cpu->regs->eflags;
  99         if (get_user(irq_enable, &lg->lguest_data->irq_enabled) == 0
 100             && !(irq_enable & X86_EFLAGS_IF))
 101                 eflags &= ~X86_EFLAGS_IF;
 102
 103         /* An interrupt is expected to push three things on the stack: the old
 104          * "eflags" word, the old code segment, and the old instruction
 105          * pointer. */
 106         push_guest_stack(lg, &gstack, eflags);
 107         push_guest_stack(lg, &gstack, cpu->regs->cs);
 108         push_guest_stack(lg, &gstack, cpu->regs->eip);
 109
 110         /* For the six traps which supply an error code, we push that, too. */
 111         if (has_err)
 112                 push_guest_stack(lg, &gstack, cpu->regs->errcode);
 113
 114         /* Now we've pushed all the old state, we change the stack, the code
 115          * segment and the address to execute. */
 116         cpu->regs->ss = ss;
 117         cpu->regs->esp = virtstack + (gstack - origstack);
 118         cpu->regs->cs = (__KERNEL_CS|GUEST_PL);
 119         cpu->regs->eip = idt_address(lo, hi);
 120
 121         /* There are two kinds of interrupt handlers: 0xE is an "interrupt
 122          * gate" which expects interrupts to be disabled on entry. */
 123         if (idt_type(lo, hi) == 0xE)
 124                 if (put_user(0, &lg->lguest_data->irq_enabled))
 125                         kill_guest(lg, "Disabling interrupts");
 126 }
 127
 128 /*H:205
 129  * Virtual Interrupts.
 130  *
 131  * maybe_do_interrupt() gets called before every entry to the Guest, to see if
 132  * we should divert the Guest to running an interrupt handler. */
 133 void maybe_do_interrupt(struct lg_cpu *cpu)
 134 {
 135         unsigned int irq;
 136         struct lguest *lg = cpu->lg;
 137         DECLARE_BITMAP(blk, LGUEST_IRQS);
 138         struct desc_struct *idt;
 139
 140         /* If the Guest hasn't even initialized yet, we can do nothing. */
 141         if (!lg->lguest_data)
 142                 return;
 143
 144         /* Take our "irqs_pending" array and remove any interrupts the Guest
 145          * wants blocked: the result ends up in "blk". */
 146         if (copy_from_user(&blk, lg->lguest_data->blocked_interrupts,
 147                            sizeof(blk)))
 148                 return;
 149
 150         bitmap_andnot(blk, cpu->irqs_pending, blk, LGUEST_IRQS);
 151
 152         /* Find the first interrupt. */
 153         irq = find_first_bit(blk, LGUEST_IRQS);
 154         /* None?  Nothing to do */
 155         if (irq >= LGUEST_IRQS)
 156                 return;
 157
 158         /* They may be in the middle of an iret, where they asked us never to
 159          * deliver interrupts. */
 160         if (cpu->regs->eip >= lg->noirq_start && cpu->regs->eip < lg->noirq_end)
 161                 return;
 162
 163         /* If they're halted, interrupts restart them. */
 164         if (cpu->halted) {
 165                 /* Re-enable interrupts. */
 166                 if (put_user(X86_EFLAGS_IF, &lg->lguest_data->irq_enabled))
 167                         kill_guest(lg, "Re-enabling interrupts");
 168                 cpu->halted = 0;
 169         } else {
 170                 /* Otherwise we check if they have interrupts disabled. */
 171                 u32 irq_enabled;
 172                 if (get_user(irq_enabled, &lg->lguest_data->irq_enabled))
 173                         irq_enabled = 0;
 174                 if (!irq_enabled)
 175                         return;
 176         }
 177
 178         /* Look at the IDT entry the Guest gave us for this interrupt.  The
 179          * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
 180          * over them. */
 181         idt = &cpu->arch.idt[FIRST_EXTERNAL_VECTOR+irq];
 182         /* If they don't have a handler (yet?), we just ignore it */
 183         if (idt_present(idt->a, idt->b)) {
 184                 /* OK, mark it no longer pending and deliver it. */
 185                 clear_bit(irq, cpu->irqs_pending);
 186                 /* set_guest_interrupt() takes the interrupt descriptor and a
 187                  * flag to say whether this interrupt pushes an error code onto
 188                  * the stack as well: virtual interrupts never do. */
 189                 set_guest_interrupt(cpu, idt->a, idt->b, 0);
 190         }
 191
 192         /* Every time we deliver an interrupt, we update the timestamp in the
 193          * Guest's lguest_data struct.  It would be better for the Guest if we
 194          * did this more often, but it can actually be quite slow: doing it
 195          * here is a compromise which means at least it gets updated every
 196          * timer interrupt. */
 197         write_timestamp(lg);
 198 }
 199 /*:*/
 200
 201 /* Linux uses trap 128 for system calls.  Plan9 uses 64, and Ron Minnich sent
 202  * me a patch, so we support that too.  It'd be a big step for lguest if half
 203  * the Plan 9 user base were to start using it.
 204  *
 205  * Actually now I think of it, it's possible that Ron *is* half the Plan 9
 206  * userbase.  Oh well. */
 207 static bool could_be_syscall(unsigned int num)
 208 {
 209         /* Normal Linux SYSCALL_VECTOR or reserved vector? */
 210         return num == SYSCALL_VECTOR || num == syscall_vector;
 211 }
 212
 213 /* The syscall vector it wants must be unused by Host. */
 214 bool check_syscall_vector(struct lguest *lg)
 215 {
 216         u32 vector;
 217
 218         if (get_user(vector, &lg->lguest_data->syscall_vec))
 219                 return false;
 220
 221         return could_be_syscall(vector);
 222 }
 223
 224 int init_interrupts(void)
 225 {
 226         /* If they want some strange system call vector, reserve it now */
 227         if (syscall_vector != SYSCALL_VECTOR
 228             && test_and_set_bit(syscall_vector, used_vectors)) {
 229                 printk("lg: couldn't reserve syscall %u\n", syscall_vector);
 230                 return -EBUSY;
 231         }
 232         return 0;
 233 }
 234
 235 void free_interrupts(void)
 236 {
 237         if (syscall_vector != SYSCALL_VECTOR)
 238                 clear_bit(syscall_vector, used_vectors);
 239 }
 240
 241 /*H:220 Now we've got the routines to deliver interrupts, delivering traps
 242  * like page fault is easy.  The only trick is that Intel decided that some
 243  * traps should have error codes: */
 244 static int has_err(unsigned int trap)
 245 {
 246         return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
 247 }
 248
 249 /* deliver_trap() returns true if it could deliver the trap. */
 250 int deliver_trap(struct lg_cpu *cpu, unsigned int num)
 251 {
 252         /* Trap numbers are always 8 bit, but we set an impossible trap number
 253          * for traps inside the Switcher, so check that here. */
 254         if (num >= ARRAY_SIZE(cpu->arch.idt))
 255                 return 0;
 256
 257         /* Early on the Guest hasn't set the IDT entries (or maybe it put a
 258          * bogus one in): if we fail here, the Guest will be killed. */
 259         if (!idt_present(cpu->arch.idt[num].a, cpu->arch.idt[num].b))
 260                 return 0;
 261         set_guest_interrupt(cpu, cpu->arch.idt[num].a,
 262                             cpu->arch.idt[num].b, has_err(num));
 263         return 1;
 264 }
 265
 266 /*H:250 Here's the hard part: returning to the Host every time a trap happens
 267  * and then calling deliver_trap() and re-entering the Guest is slow.
 268  * Particularly because Guest userspace system calls are traps (usually trap
 269  * 128).
 270  *
 271  * So we'd like to set up the IDT to tell the CPU to deliver traps directly
 272  * into the Guest.  This is possible, but the complexities cause the size of
 273  * this file to double!  However, 150 lines of code is worth writing for taking
 274  * system calls down from 1750ns to 270ns.  Plus, if lguest didn't do it, all
 275  * the other hypervisors would beat it up at lunchtime.
 276  *
 277  * This routine indicates if a particular trap number could be delivered
 278  * directly. */
 279 static int direct_trap(unsigned int num)
 280 {
 281         /* Hardware interrupts don't go to the Guest at all (except system
 282          * call). */
 283         if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
 284                 return 0;
 285
 286         /* The Host needs to see page faults (for shadow paging and to save the
 287          * fault address), general protection faults (in/out emulation) and
 288          * device not available (TS handling), and of course, the hypercall
 289          * trap. */
 290         return num != 14 && num != 13 && num != 7 && num != LGUEST_TRAP_ENTRY;
 291 }
 292 /*:*/
 293
 294 /*M:005 The Guest has the ability to turn its interrupt gates into trap gates,
 295  * if it is careful.  The Host will let trap gates can go directly to the
 296  * Guest, but the Guest needs the interrupts atomically disabled for an
 297  * interrupt gate.  It can do this by pointing the trap gate at instructions
 298  * within noirq_start and noirq_end, where it can safely disable interrupts. */
 299
 300 /*M:006 The Guests do not use the sysenter (fast system call) instruction,
 301  * because it's hardcoded to enter privilege level 0 and so can't go direct.
 302  * It's about twice as fast as the older "int 0x80" system call, so it might
 303  * still be worthwhile to handle it in the Switcher and lcall down to the
 304  * Guest.  The sysenter semantics are hairy tho: search for that keyword in
 305  * entry.S :*/
 306
 307 /*H:260 When we make traps go directly into the Guest, we need to make sure
 308  * the kernel stack is valid (ie. mapped in the page tables).  Otherwise, the
 309  * CPU trying to deliver the trap will fault while trying to push the interrupt
 310  * words on the stack: this is called a double fault, and it forces us to kill
 311  * the Guest.
 312  *
 313  * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */
 314 void pin_stack_pages(struct lg_cpu *cpu)
 315 {
 316         unsigned int i;
 317
 318         struct lguest *lg = cpu->lg;
 319         /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or
 320          * two pages of stack space. */
 321         for (i = 0; i < lg->stack_pages; i++)
 322                 /* The stack grows *upwards*, so the address we're given is the
 323                  * start of the page after the kernel stack.  Subtract one to
 324                  * get back onto the first stack page, and keep subtracting to
 325                  * get to the rest of the stack pages. */
 326                 pin_page(lg, cpu->esp1 - 1 - i * PAGE_SIZE);
 327 }
 328
 329 /* Direct traps also mean that we need to know whenever the Guest wants to use
 330  * a different kernel stack, so we can change the IDT entries to use that
 331  * stack.  The IDT entries expect a virtual address, so unlike most addresses
 332  * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
 333  * physical.
 334  *
 335  * In Linux each process has its own kernel stack, so this happens a lot: we
 336  * change stacks on each context switch. */
 337 void guest_set_stack(struct lg_cpu *cpu, u32 seg, u32 esp, unsigned int pages)
 338 {
 339         /* You are not allowed have a stack segment with privilege level 0: bad
 340          * Guest! */
 341         if ((seg & 0x3) != GUEST_PL)
 342                 kill_guest(cpu->lg, "bad stack segment %i", seg);
 343         /* We only expect one or two stack pages. */
 344         if (pages > 2)
 345                 kill_guest(cpu->lg, "bad stack pages %u", pages);
 346         /* Save where the stack is, and how many pages */
 347         cpu->ss1 = seg;
 348         cpu->esp1 = esp;
 349         cpu->lg->stack_pages = pages;
 350         /* Make sure the new stack pages are mapped */
 351         pin_stack_pages(cpu);
 352 }
 353
 354 /* All this reference to mapping stacks leads us neatly into the other complex
 355  * part of the Host: page table handling. */
 356
 357 /*H:235 This is the routine which actually checks the Guest's IDT entry and
 358  * transfers it into the entry in "struct lguest": */
 359 static void set_trap(struct lguest *lg, struct desc_struct *trap,
 360                      unsigned int num, u32 lo, u32 hi)
 361 {
 362         u8 type = idt_type(lo, hi);
 363
 364         /* We zero-out a not-present entry */
 365         if (!idt_present(lo, hi)) {
 366                 trap->a = trap->b = 0;
 367                 return;
 368         }
 369
 370         /* We only support interrupt and trap gates. */
 371         if (type != 0xE && type != 0xF)
 372                 kill_guest(lg, "bad IDT type %i", type);
 373
 374         /* We only copy the handler address, present bit, privilege level and
 375          * type.  The privilege level controls where the trap can be triggered
 376          * manually with an "int" instruction.  This is usually GUEST_PL,
 377          * except for system calls which userspace can use. */
 378         trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
 379         trap->b = (hi&0xFFFFEF00);
 380 }
 381
 382 /*H:230 While we're here, dealing with delivering traps and interrupts to the
 383  * Guest, we might as well complete the picture: how the Guest tells us where
 384  * it wants them to go.  This would be simple, except making traps fast
 385  * requires some tricks.
 386  *
 387  * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
 388  * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */
 389 void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int num, u32 lo, u32 hi)
 390 {
 391         /* Guest never handles: NMI, doublefault, spurious interrupt or
 392          * hypercall.  We ignore when it tries to set them. */
 393         if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
 394                 return;
 395
 396         /* Mark the IDT as changed: next time the Guest runs we'll know we have
 397          * to copy this again. */
 398         cpu->lg->changed |= CHANGED_IDT;
 399
 400         /* Check that the Guest doesn't try to step outside the bounds. */
 401         if (num >= ARRAY_SIZE(cpu->arch.idt))
 402                 kill_guest(cpu->lg, "Setting idt entry %u", num);
 403         else
 404                 set_trap(cpu->lg, &cpu->arch.idt[num], num, lo, hi);
 405 }
 406
 407 /* The default entry for each interrupt points into the Switcher routines which
 408  * simply return to the Host.  The run_guest() loop will then call
 409  * deliver_trap() to bounce it back into the Guest. */
 410 static void default_idt_entry(struct desc_struct *idt,
 411                               int trap,
 412                               const unsigned long handler)
 413 {
 414         /* A present interrupt gate. */
 415         u32 flags = 0x8e00;
 416
 417         /* Set the privilege level on the entry for the hypercall: this allows
 418          * the Guest to use the "int" instruction to trigger it. */
 419         if (trap == LGUEST_TRAP_ENTRY)
 420                 flags |= (GUEST_PL << 13);
 421
 422         /* Now pack it into the IDT entry in its weird format. */
 423         idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
 424         idt->b = (handler&0xFFFF0000) | flags;
 425 }
 426
 427 /* When the Guest first starts, we put default entries into the IDT. */
 428 void setup_default_idt_entries(struct lguest_ro_state *state,
 429                                const unsigned long *def)
 430 {
 431         unsigned int i;
 432
 433         for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
 434                 default_idt_entry(&state->guest_idt[i], i, def[i]);
 435 }
 436
 437 /*H:240 We don't use the IDT entries in the "struct lguest" directly, instead
 438  * we copy them into the IDT which we've set up for Guests on this CPU, just
 439  * before we run the Guest.  This routine does that copy. */
 440 void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt,
 441                 const unsigned long *def)
 442 {
 443         unsigned int i;
 444
 445         /* We can simply copy the direct traps, otherwise we use the default
 446          * ones in the Switcher: they will return to the Host. */
 447         for (i = 0; i < ARRAY_SIZE(cpu->arch.idt); i++) {
 448                 /* If no Guest can ever override this trap, leave it alone. */
 449                 if (!direct_trap(i))
 450                         continue;
 451
 452                 /* Only trap gates (type 15) can go direct to the Guest.
 453                  * Interrupt gates (type 14) disable interrupts as they are
 454                  * entered, which we never let the Guest do.  Not present
 455                  * entries (type 0x0) also can't go direct, of course. */
 456                 if (idt_type(cpu->arch.idt[i].a, cpu->arch.idt[i].b) == 0xF)
 457                         idt[i] = cpu->arch.idt[i];
 458                 else
 459                         /* Reset it to the default. */
 460                         default_idt_entry(&idt[i], i, def[i]);
 461         }
 462 }
 463
 464 /*H:200
 465  * The Guest Clock.
 466  *
 467  * There are two sources of virtual interrupts.  We saw one in lguest_user.c:
 468  * the Launcher sending interrupts for virtual devices.  The other is the Guest
 469  * timer interrupt.
 470  *
 471  * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
 472  * the next timer interrupt (in nanoseconds).  We use the high-resolution timer
 473  * infrastructure to set a callback at that time.
 474  *
 475  * 0 means "turn off the clock". */
 476 void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta)
 477 {
 478         ktime_t expires;
 479
 480         if (unlikely(delta == 0)) {
 481                 /* Clock event device is shutting down. */
 482                 hrtimer_cancel(&cpu->hrt);
 483                 return;
 484         }
 485
 486         /* We use wallclock time here, so the Guest might not be running for
 487          * all the time between now and the timer interrupt it asked for.  This
 488          * is almost always the right thing to do. */
 489         expires = ktime_add_ns(ktime_get_real(), delta);
 490         hrtimer_start(&cpu->hrt, expires, HRTIMER_MODE_ABS);
 491 }
 492
 493 /* This is the function called when the Guest's timer expires. */
 494 static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
 495 {
 496         struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt);
 497
 498         /* Remember the first interrupt is the timer interrupt. */
 499         set_bit(0, cpu->irqs_pending);
 500         /* If the Guest is actually stopped, we need to wake it up. */
 501         if (cpu->halted)
 502                 wake_up_process(cpu->tsk);
 503         return HRTIMER_NORESTART;
 504 }
 505
 506 /* This sets up the timer for this Guest. */
 507 void init_clockdev(struct lg_cpu *cpu)
 508 {
 509         hrtimer_init(&cpu->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
 510         cpu->hrt.function = clockdev_fn;
 511 }