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