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 bool 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,
  64                                 bool has_err)
  65 {
  66         unsigned long gstack, origstack;
  67         u32 eflags, ss, irq_enable;
  68         unsigned long virtstack;
  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(cpu, 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(cpu, &gstack, cpu->regs->ss);
  85                 push_guest_stack(cpu, &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(cpu, 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, &cpu->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(cpu, &gstack, eflags);
 107         push_guest_stack(cpu, &gstack, cpu->regs->cs);
 108         push_guest_stack(cpu, &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(cpu, &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, &cpu->lg->lguest_data->irq_enabled))
 125                         kill_guest(cpu, "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         DECLARE_BITMAP(blk, LGUEST_IRQS);
 137         struct desc_struct *idt;
 138
 139         /* If the Guest hasn't even initialized yet, we can do nothing. */
 140         if (!cpu->lg->lguest_data)
 141                 return;
 142
 143         /* Take our "irqs_pending" array and remove any interrupts the Guest
 144          * wants blocked: the result ends up in "blk". */
 145         if (copy_from_user(&blk, cpu->lg->lguest_data->blocked_interrupts,
 146                            sizeof(blk)))
 147                 return;
 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, false);
 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                 if (test_bit(syscall_vector, used_vectors) ||
 228                     vector_used_by_percpu_irq(syscall_vector)) {
 229                         printk(KERN_ERR "lg: couldn't reserve syscall %u\n",
 230                                  syscall_vector);
 231                         return -EBUSY;
 232                 }
 233                 set_bit(syscall_vector, used_vectors);
 234         }
 235
 236         return 0;
 237 }
 238
 239 void free_interrupts(void)
 240 {
 241         if (syscall_vector != SYSCALL_VECTOR)
 242                 clear_bit(syscall_vector, used_vectors);
 243 }
 244
 245 /*H:220 Now we've got the routines to deliver interrupts, delivering traps like
 246  * page fault is easy.  The only trick is that Intel decided that some traps
 247  * should have error codes: */
 248 static bool has_err(unsigned int trap)
 249 {
 250         return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
 251 }
 252
 253 /* deliver_trap() returns true if it could deliver the trap. */
 254 bool deliver_trap(struct lg_cpu *cpu, unsigned int num)
 255 {
 256         /* Trap numbers are always 8 bit, but we set an impossible trap number
 257          * for traps inside the Switcher, so check that here. */
 258         if (num >= ARRAY_SIZE(cpu->arch.idt))
 259                 return false;
 260
 261         /* Early on the Guest hasn't set the IDT entries (or maybe it put a
 262          * bogus one in): if we fail here, the Guest will be killed. */
 263         if (!idt_present(cpu->arch.idt[num].a, cpu->arch.idt[num].b))
 264                 return false;
 265         set_guest_interrupt(cpu, cpu->arch.idt[num].a,
 266                             cpu->arch.idt[num].b, has_err(num));
 267         return true;
 268 }
 269
 270 /*H:250 Here's the hard part: returning to the Host every time a trap happens
 271  * and then calling deliver_trap() and re-entering the Guest is slow.
 272  * Particularly because Guest userspace system calls are traps (usually trap
 273  * 128).
 274  *
 275  * So we'd like to set up the IDT to tell the CPU to deliver traps directly
 276  * into the Guest.  This is possible, but the complexities cause the size of
 277  * this file to double!  However, 150 lines of code is worth writing for taking
 278  * system calls down from 1750ns to 270ns.  Plus, if lguest didn't do it, all
 279  * the other hypervisors would beat it up at lunchtime.
 280  *
 281  * This routine indicates if a particular trap number could be delivered
 282  * directly. */
 283 static bool direct_trap(unsigned int num)
 284 {
 285         /* Hardware interrupts don't go to the Guest at all (except system
 286          * call). */
 287         if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
 288                 return false;
 289
 290         /* The Host needs to see page faults (for shadow paging and to save the
 291          * fault address), general protection faults (in/out emulation) and
 292          * device not available (TS handling), invalid opcode fault (kvm hcall),
 293          * and of course, the hypercall trap. */
 294         return num != 14 && num != 13 && num != 7 &&
 295                         num != 6 && num != LGUEST_TRAP_ENTRY;
 296 }
 297 /*:*/
 298
 299 /*M:005 The Guest has the ability to turn its interrupt gates into trap gates,
 300  * if it is careful.  The Host will let trap gates can go directly to the
 301  * Guest, but the Guest needs the interrupts atomically disabled for an
 302  * interrupt gate.  It can do this by pointing the trap gate at instructions
 303  * within noirq_start and noirq_end, where it can safely disable interrupts. */
 304
 305 /*M:006 The Guests do not use the sysenter (fast system call) instruction,
 306  * because it's hardcoded to enter privilege level 0 and so can't go direct.
 307  * It's about twice as fast as the older "int 0x80" system call, so it might
 308  * still be worthwhile to handle it in the Switcher and lcall down to the
 309  * Guest.  The sysenter semantics are hairy tho: search for that keyword in
 310  * entry.S :*/
 311
 312 /*H:260 When we make traps go directly into the Guest, we need to make sure
 313  * the kernel stack is valid (ie. mapped in the page tables).  Otherwise, the
 314  * CPU trying to deliver the trap will fault while trying to push the interrupt
 315  * words on the stack: this is called a double fault, and it forces us to kill
 316  * the Guest.
 317  *
 318  * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */
 319 void pin_stack_pages(struct lg_cpu *cpu)
 320 {
 321         unsigned int i;
 322
 323         /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or
 324          * two pages of stack space. */
 325         for (i = 0; i < cpu->lg->stack_pages; i++)
 326                 /* The stack grows *upwards*, so the address we're given is the
 327                  * start of the page after the kernel stack.  Subtract one to
 328                  * get back onto the first stack page, and keep subtracting to
 329                  * get to the rest of the stack pages. */
 330                 pin_page(cpu, cpu->esp1 - 1 - i * PAGE_SIZE);
 331 }
 332
 333 /* Direct traps also mean that we need to know whenever the Guest wants to use
 334  * a different kernel stack, so we can change the IDT entries to use that
 335  * stack.  The IDT entries expect a virtual address, so unlike most addresses
 336  * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
 337  * physical.
 338  *
 339  * In Linux each process has its own kernel stack, so this happens a lot: we
 340  * change stacks on each context switch. */
 341 void guest_set_stack(struct lg_cpu *cpu, u32 seg, u32 esp, unsigned int pages)
 342 {
 343         /* You are not allowed have a stack segment with privilege level 0: bad
 344          * Guest! */
 345         if ((seg & 0x3) != GUEST_PL)
 346                 kill_guest(cpu, "bad stack segment %i", seg);
 347         /* We only expect one or two stack pages. */
 348         if (pages > 2)
 349                 kill_guest(cpu, "bad stack pages %u", pages);
 350         /* Save where the stack is, and how many pages */
 351         cpu->ss1 = seg;
 352         cpu->esp1 = esp;
 353         cpu->lg->stack_pages = pages;
 354         /* Make sure the new stack pages are mapped */
 355         pin_stack_pages(cpu);
 356 }
 357
 358 /* All this reference to mapping stacks leads us neatly into the other complex
 359  * part of the Host: page table handling. */
 360
 361 /*H:235 This is the routine which actually checks the Guest's IDT entry and
 362  * transfers it into the entry in "struct lguest": */
 363 static void set_trap(struct lg_cpu *cpu, struct desc_struct *trap,
 364                      unsigned int num, u32 lo, u32 hi)
 365 {
 366         u8 type = idt_type(lo, hi);
 367
 368         /* We zero-out a not-present entry */
 369         if (!idt_present(lo, hi)) {
 370                 trap->a = trap->b = 0;
 371                 return;
 372         }
 373
 374         /* We only support interrupt and trap gates. */
 375         if (type != 0xE && type != 0xF)
 376                 kill_guest(cpu, "bad IDT type %i", type);
 377
 378         /* We only copy the handler address, present bit, privilege level and
 379          * type.  The privilege level controls where the trap can be triggered
 380          * manually with an "int" instruction.  This is usually GUEST_PL,
 381          * except for system calls which userspace can use. */
 382         trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
 383         trap->b = (hi&0xFFFFEF00);
 384 }
 385
 386 /*H:230 While we're here, dealing with delivering traps and interrupts to the
 387  * Guest, we might as well complete the picture: how the Guest tells us where
 388  * it wants them to go.  This would be simple, except making traps fast
 389  * requires some tricks.
 390  *
 391  * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
 392  * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */
 393 void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int num, u32 lo, u32 hi)
 394 {
 395         /* Guest never handles: NMI, doublefault, spurious interrupt or
 396          * hypercall.  We ignore when it tries to set them. */
 397         if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
 398                 return;
 399
 400         /* Mark the IDT as changed: next time the Guest runs we'll know we have
 401          * to copy this again. */
 402         cpu->changed |= CHANGED_IDT;
 403
 404         /* Check that the Guest doesn't try to step outside the bounds. */
 405         if (num >= ARRAY_SIZE(cpu->arch.idt))
 406                 kill_guest(cpu, "Setting idt entry %u", num);
 407         else
 408                 set_trap(cpu, &cpu->arch.idt[num], num, lo, hi);
 409 }
 410
 411 /* The default entry for each interrupt points into the Switcher routines which
 412  * simply return to the Host.  The run_guest() loop will then call
 413  * deliver_trap() to bounce it back into the Guest. */
 414 static void default_idt_entry(struct desc_struct *idt,
 415                               int trap,
 416                               const unsigned long handler,
 417                               const struct desc_struct *base)
 418 {
 419         /* A present interrupt gate. */
 420         u32 flags = 0x8e00;
 421
 422         /* Set the privilege level on the entry for the hypercall: this allows
 423          * the Guest to use the "int" instruction to trigger it. */
 424         if (trap == LGUEST_TRAP_ENTRY)
 425                 flags |= (GUEST_PL << 13);
 426         else if (base)
 427                 /* Copy priv. level from what Guest asked for.  This allows
 428                  * debug (int 3) traps from Guest userspace, for example. */
 429                 flags |= (base->b & 0x6000);
 430
 431         /* Now pack it into the IDT entry in its weird format. */
 432         idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
 433         idt->b = (handler&0xFFFF0000) | flags;
 434 }
 435
 436 /* When the Guest first starts, we put default entries into the IDT. */
 437 void setup_default_idt_entries(struct lguest_ro_state *state,
 438                                const unsigned long *def)
 439 {
 440         unsigned int i;
 441
 442         for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
 443                 default_idt_entry(&state->guest_idt[i], i, def[i], NULL);
 444 }
 445
 446 /*H:240 We don't use the IDT entries in the "struct lguest" directly, instead
 447  * we copy them into the IDT which we've set up for Guests on this CPU, just
 448  * before we run the Guest.  This routine does that copy. */
 449 void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt,
 450                 const unsigned long *def)
 451 {
 452         unsigned int i;
 453
 454         /* We can simply copy the direct traps, otherwise we use the default
 455          * ones in the Switcher: they will return to the Host. */
 456         for (i = 0; i < ARRAY_SIZE(cpu->arch.idt); i++) {
 457                 const struct desc_struct *gidt = &cpu->arch.idt[i];
 458
 459                 /* If no Guest can ever override this trap, leave it alone. */
 460                 if (!direct_trap(i))
 461                         continue;
 462
 463                 /* Only trap gates (type 15) can go direct to the Guest.
 464                  * Interrupt gates (type 14) disable interrupts as they are
 465                  * entered, which we never let the Guest do.  Not present
 466                  * entries (type 0x0) also can't go direct, of course.
 467                  *
 468                  * If it can't go direct, we still need to copy the priv. level:
 469                  * they might want to give userspace access to a software
 470                  * interrupt. */
 471                 if (idt_type(gidt->a, gidt->b) == 0xF)
 472                         idt[i] = *gidt;
 473                 else
 474                         default_idt_entry(&idt[i], i, def[i], gidt);
 475         }
 476 }
 477
 478 /*H:200
 479  * The Guest Clock.
 480  *
 481  * There are two sources of virtual interrupts.  We saw one in lguest_user.c:
 482  * the Launcher sending interrupts for virtual devices.  The other is the Guest
 483  * timer interrupt.
 484  *
 485  * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
 486  * the next timer interrupt (in nanoseconds).  We use the high-resolution timer
 487  * infrastructure to set a callback at that time.
 488  *
 489  * 0 means "turn off the clock". */
 490 void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta)
 491 {
 492         ktime_t expires;
 493
 494         if (unlikely(delta == 0)) {
 495                 /* Clock event device is shutting down. */
 496                 hrtimer_cancel(&cpu->hrt);
 497                 return;
 498         }
 499
 500         /* We use wallclock time here, so the Guest might not be running for
 501          * all the time between now and the timer interrupt it asked for.  This
 502          * is almost always the right thing to do. */
 503         expires = ktime_add_ns(ktime_get_real(), delta);
 504         hrtimer_start(&cpu->hrt, expires, HRTIMER_MODE_ABS);
 505 }
 506
 507 /* This is the function called when the Guest's timer expires. */
 508 static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
 509 {
 510         struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt);
 511
 512         /* Remember the first interrupt is the timer interrupt. */
 513         set_bit(0, cpu->irqs_pending);
 514         /* If the Guest is actually stopped, we need to wake it up. */
 515         if (cpu->halted)
 516                 wake_up_process(cpu->tsk);
 517         return HRTIMER_NORESTART;
 518 }
 519
 520 /* This sets up the timer for this Guest. */
 521 void init_clockdev(struct lg_cpu *cpu)
 522 {
 523         hrtimer_init(&cpu->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
 524         cpu->hrt.function = clockdev_fn;
 525 }