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 "lg.h"
  16
  17 /* The address of the interrupt handler is split into two bits: */
  18 static unsigned long idt_address(u32 lo, u32 hi)
  19 {
  20         return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
  21 }
  22
  23 /* The "type" of the interrupt handler is a 4 bit field: we only support a
  24  * couple of types. */
  25 static int idt_type(u32 lo, u32 hi)
  26 {
  27         return (hi >> 8) & 0xF;
  28 }
  29
  30 /* An IDT entry can't be used unless the "present" bit is set. */
  31 static int idt_present(u32 lo, u32 hi)
  32 {
  33         return (hi & 0x8000);
  34 }
  35
  36 /* We need a helper to "push" a value onto the Guest's stack, since that's a
  37  * big part of what delivering an interrupt does. */
  38 static void push_guest_stack(struct lguest *lg, unsigned long *gstack, u32 val)
  39 {
  40         /* Stack grows upwards: move stack then write value. */
  41         *gstack -= 4;
  42         lgwrite_u32(lg, *gstack, val);
  43 }
  44
  45 /*H:210 The set_guest_interrupt() routine actually delivers the interrupt or
  46  * trap.  The mechanics of delivering traps and interrupts to the Guest are the
  47  * same, except some traps have an "error code" which gets pushed onto the
  48  * stack as well: the caller tells us if this is one.
  49  *
  50  * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
  51  * interrupt or trap.  It's split into two parts for traditional reasons: gcc
  52  * on i386 used to be frightened by 64 bit numbers.
  53  *
  54  * We set up the stack just like the CPU does for a real interrupt, so it's
  55  * identical for the Guest (and the standard "iret" instruction will undo
  56  * it). */
  57 static void set_guest_interrupt(struct lguest *lg, u32 lo, u32 hi, int has_err)
  58 {
  59         unsigned long gstack;
  60         u32 eflags, ss, irq_enable;
  61
  62         /* There are two cases for interrupts: one where the Guest is already
  63          * in the kernel, and a more complex one where the Guest is in
  64          * userspace.  We check the privilege level to find out. */
  65         if ((lg->regs->ss&0x3) != GUEST_PL) {
  66                 /* The Guest told us their kernel stack with the SET_STACK
  67                  * hypercall: both the virtual address and the segment */
  68                 gstack = guest_pa(lg, lg->esp1);
  69                 ss = lg->ss1;
  70                 /* We push the old stack segment and pointer onto the new
  71                  * stack: when the Guest does an "iret" back from the interrupt
  72                  * handler the CPU will notice they're dropping privilege
  73                  * levels and expect these here. */
  74                 push_guest_stack(lg, &gstack, lg->regs->ss);
  75                 push_guest_stack(lg, &gstack, lg->regs->esp);
  76         } else {
  77                 /* We're staying on the same Guest (kernel) stack. */
  78                 gstack = guest_pa(lg, lg->regs->esp);
  79                 ss = lg->regs->ss;
  80         }
  81
  82         /* Remember that we never let the Guest actually disable interrupts, so
  83          * the "Interrupt Flag" bit is always set.  We copy that bit from the
  84          * Guest's "irq_enabled" field into the eflags word: the Guest copies
  85          * it back in "lguest_iret". */
  86         eflags = lg->regs->eflags;
  87         if (get_user(irq_enable, &lg->lguest_data->irq_enabled) == 0
  88             && !(irq_enable & X86_EFLAGS_IF))
  89                 eflags &= ~X86_EFLAGS_IF;
  90
  91         /* An interrupt is expected to push three things on the stack: the old
  92          * "eflags" word, the old code segment, and the old instruction
  93          * pointer. */
  94         push_guest_stack(lg, &gstack, eflags);
  95         push_guest_stack(lg, &gstack, lg->regs->cs);
  96         push_guest_stack(lg, &gstack, lg->regs->eip);
  97
  98         /* For the six traps which supply an error code, we push that, too. */
  99         if (has_err)
 100                 push_guest_stack(lg, &gstack, lg->regs->errcode);
 101
 102         /* Now we've pushed all the old state, we change the stack, the code
 103          * segment and the address to execute. */
 104         lg->regs->ss = ss;
 105         lg->regs->esp = gstack + lg->page_offset;
 106         lg->regs->cs = (__KERNEL_CS|GUEST_PL);
 107         lg->regs->eip = idt_address(lo, hi);
 108
 109         /* There are two kinds of interrupt handlers: 0xE is an "interrupt
 110          * gate" which expects interrupts to be disabled on entry. */
 111         if (idt_type(lo, hi) == 0xE)
 112                 if (put_user(0, &lg->lguest_data->irq_enabled))
 113                         kill_guest(lg, "Disabling interrupts");
 114 }
 115
 116 /*H:200
 117  * Virtual Interrupts.
 118  *
 119  * maybe_do_interrupt() gets called before every entry to the Guest, to see if
 120  * we should divert the Guest to running an interrupt handler. */
 121 void maybe_do_interrupt(struct lguest *lg)
 122 {
 123         unsigned int irq;
 124         DECLARE_BITMAP(blk, LGUEST_IRQS);
 125         struct desc_struct *idt;
 126
 127         /* If the Guest hasn't even initialized yet, we can do nothing. */
 128         if (!lg->lguest_data)
 129                 return;
 130
 131         /* Take our "irqs_pending" array and remove any interrupts the Guest
 132          * wants blocked: the result ends up in "blk". */
 133         if (copy_from_user(&blk, lg->lguest_data->blocked_interrupts,
 134                            sizeof(blk)))
 135                 return;
 136
 137         bitmap_andnot(blk, lg->irqs_pending, blk, LGUEST_IRQS);
 138
 139         /* Find the first interrupt. */
 140         irq = find_first_bit(blk, LGUEST_IRQS);
 141         /* None?  Nothing to do */
 142         if (irq >= LGUEST_IRQS)
 143                 return;
 144
 145         /* They may be in the middle of an iret, where they asked us never to
 146          * deliver interrupts. */
 147         if (lg->regs->eip >= lg->noirq_start && lg->regs->eip < lg->noirq_end)
 148                 return;
 149
 150         /* If they're halted, interrupts restart them. */
 151         if (lg->halted) {
 152                 /* Re-enable interrupts. */
 153                 if (put_user(X86_EFLAGS_IF, &lg->lguest_data->irq_enabled))
 154                         kill_guest(lg, "Re-enabling interrupts");
 155                 lg->halted = 0;
 156         } else {
 157                 /* Otherwise we check if they have interrupts disabled. */
 158                 u32 irq_enabled;
 159                 if (get_user(irq_enabled, &lg->lguest_data->irq_enabled))
 160                         irq_enabled = 0;
 161                 if (!irq_enabled)
 162                         return;
 163         }
 164
 165         /* Look at the IDT entry the Guest gave us for this interrupt.  The
 166          * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
 167          * over them. */
 168         idt = &lg->idt[FIRST_EXTERNAL_VECTOR+irq];
 169         /* If they don't have a handler (yet?), we just ignore it */
 170         if (idt_present(idt->a, idt->b)) {
 171                 /* OK, mark it no longer pending and deliver it. */
 172                 clear_bit(irq, lg->irqs_pending);
 173                 /* set_guest_interrupt() takes the interrupt descriptor and a
 174                  * flag to say whether this interrupt pushes an error code onto
 175                  * the stack as well: virtual interrupts never do. */
 176                 set_guest_interrupt(lg, idt->a, idt->b, 0);
 177         }
 178
 179         /* Every time we deliver an interrupt, we update the timestamp in the
 180          * Guest's lguest_data struct.  It would be better for the Guest if we
 181          * did this more often, but it can actually be quite slow: doing it
 182          * here is a compromise which means at least it gets updated every
 183          * timer interrupt. */
 184         write_timestamp(lg);
 185 }
 186
 187 /*H:220 Now we've got the routines to deliver interrupts, delivering traps
 188  * like page fault is easy.  The only trick is that Intel decided that some
 189  * traps should have error codes: */
 190 static int has_err(unsigned int trap)
 191 {
 192         return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
 193 }
 194
 195 /* deliver_trap() returns true if it could deliver the trap. */
 196 int deliver_trap(struct lguest *lg, unsigned int num)
 197 {
 198         /* Trap numbers are always 8 bit, but we set an impossible trap number
 199          * for traps inside the Switcher, so check that here. */
 200         if (num >= ARRAY_SIZE(lg->idt))
 201                 return 0;
 202
 203         /* Early on the Guest hasn't set the IDT entries (or maybe it put a
 204          * bogus one in): if we fail here, the Guest will be killed. */
 205         if (!idt_present(lg->idt[num].a, lg->idt[num].b))
 206                 return 0;
 207         set_guest_interrupt(lg, lg->idt[num].a, lg->idt[num].b, has_err(num));
 208         return 1;
 209 }
 210
 211 /*H:250 Here's the hard part: returning to the Host every time a trap happens
 212  * and then calling deliver_trap() and re-entering the Guest is slow.
 213  * Particularly because Guest userspace system calls are traps (trap 128).
 214  *
 215  * So we'd like to set up the IDT to tell the CPU to deliver traps directly
 216  * into the Guest.  This is possible, but the complexities cause the size of
 217  * this file to double!  However, 150 lines of code is worth writing for taking
 218  * system calls down from 1750ns to 270ns.  Plus, if lguest didn't do it, all
 219  * the other hypervisors would tease it.
 220  *
 221  * This routine determines if a trap can be delivered directly. */
 222 static int direct_trap(const struct lguest *lg,
 223                        const struct desc_struct *trap,
 224                        unsigned int num)
 225 {
 226         /* Hardware interrupts don't go to the Guest at all (except system
 227          * call). */
 228         if (num >= FIRST_EXTERNAL_VECTOR && num != SYSCALL_VECTOR)
 229                 return 0;
 230
 231         /* The Host needs to see page faults (for shadow paging and to save the
 232          * fault address), general protection faults (in/out emulation) and
 233          * device not available (TS handling), and of course, the hypercall
 234          * trap. */
 235         if (num == 14 || num == 13 || num == 7 || num == LGUEST_TRAP_ENTRY)
 236                 return 0;
 237
 238         /* Only trap gates (type 15) can go direct to the Guest.  Interrupt
 239          * gates (type 14) disable interrupts as they are entered, which we
 240          * never let the Guest do.  Not present entries (type 0x0) also can't
 241          * go direct, of course 8) */
 242         return idt_type(trap->a, trap->b) == 0xF;
 243 }
 244 /*:*/
 245
 246 /*M:005 The Guest has the ability to turn its interrupt gates into trap gates,
 247  * if it is careful.  The Host will let trap gates can go directly to the
 248  * Guest, but the Guest needs the interrupts atomically disabled for an
 249  * interrupt gate.  It can do this by pointing the trap gate at instructions
 250  * within noirq_start and noirq_end, where it can safely disable interrupts. */
 251
 252 /*M:006 The Guests do not use the sysenter (fast system call) instruction,
 253  * because it's hardcoded to enter privilege level 0 and so can't go direct.
 254  * It's about twice as fast as the older "int 0x80" system call, so it might
 255  * still be worthwhile to handle it in the Switcher and lcall down to the
 256  * Guest.  The sysenter semantics are hairy tho: search for that keyword in
 257  * entry.S :*/
 258
 259 /*H:260 When we make traps go directly into the Guest, we need to make sure
 260  * the kernel stack is valid (ie. mapped in the page tables).  Otherwise, the
 261  * CPU trying to deliver the trap will fault while trying to push the interrupt
 262  * words on the stack: this is called a double fault, and it forces us to kill
 263  * the Guest.
 264  *
 265  * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */
 266 void pin_stack_pages(struct lguest *lg)
 267 {
 268         unsigned int i;
 269
 270         /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or
 271          * two pages of stack space. */
 272         for (i = 0; i < lg->stack_pages; i++)
 273                 /* The stack grows *upwards*, hence the subtraction */
 274                 pin_page(lg, lg->esp1 - i * PAGE_SIZE);
 275 }
 276
 277 /* Direct traps also mean that we need to know whenever the Guest wants to use
 278  * a different kernel stack, so we can change the IDT entries to use that
 279  * stack.  The IDT entries expect a virtual address, so unlike most addresses
 280  * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
 281  * physical.
 282  *
 283  * In Linux each process has its own kernel stack, so this happens a lot: we
 284  * change stacks on each context switch. */
 285 void guest_set_stack(struct lguest *lg, u32 seg, u32 esp, unsigned int pages)
 286 {
 287         /* You are not allowd have a stack segment with privilege level 0: bad
 288          * Guest! */
 289         if ((seg & 0x3) != GUEST_PL)
 290                 kill_guest(lg, "bad stack segment %i", seg);
 291         /* We only expect one or two stack pages. */
 292         if (pages > 2)
 293                 kill_guest(lg, "bad stack pages %u", pages);
 294         /* Save where the stack is, and how many pages */
 295         lg->ss1 = seg;
 296         lg->esp1 = esp;
 297         lg->stack_pages = pages;
 298         /* Make sure the new stack pages are mapped */
 299         pin_stack_pages(lg);
 300 }
 301
 302 /* All this reference to mapping stacks leads us neatly into the other complex
 303  * part of the Host: page table handling. */
 304
 305 /*H:235 This is the routine which actually checks the Guest's IDT entry and
 306  * transfers it into our entry in "struct lguest": */
 307 static void set_trap(struct lguest *lg, struct desc_struct *trap,
 308                      unsigned int num, u32 lo, u32 hi)
 309 {
 310         u8 type = idt_type(lo, hi);
 311
 312         /* We zero-out a not-present entry */
 313         if (!idt_present(lo, hi)) {
 314                 trap->a = trap->b = 0;
 315                 return;
 316         }
 317
 318         /* We only support interrupt and trap gates. */
 319         if (type != 0xE && type != 0xF)
 320                 kill_guest(lg, "bad IDT type %i", type);
 321
 322         /* We only copy the handler address, present bit, privilege level and
 323          * type.  The privilege level controls where the trap can be triggered
 324          * manually with an "int" instruction.  This is usually GUEST_PL,
 325          * except for system calls which userspace can use. */
 326         trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
 327         trap->b = (hi&0xFFFFEF00);
 328 }
 329
 330 /*H:230 While we're here, dealing with delivering traps and interrupts to the
 331  * Guest, we might as well complete the picture: how the Guest tells us where
 332  * it wants them to go.  This would be simple, except making traps fast
 333  * requires some tricks.
 334  *
 335  * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
 336  * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */
 337 void load_guest_idt_entry(struct lguest *lg, unsigned int num, u32 lo, u32 hi)
 338 {
 339         /* Guest never handles: NMI, doublefault, spurious interrupt or
 340          * hypercall.  We ignore when it tries to set them. */
 341         if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
 342                 return;
 343
 344         /* Mark the IDT as changed: next time the Guest runs we'll know we have
 345          * to copy this again. */
 346         lg->changed |= CHANGED_IDT;
 347
 348         /* The IDT which we keep in "struct lguest" only contains 32 entries
 349          * for the traps and LGUEST_IRQS (32) entries for interrupts.  We
 350          * ignore attempts to set handlers for higher interrupt numbers, except
 351          * for the system call "interrupt" at 128: we have a special IDT entry
 352          * for that. */
 353         if (num < ARRAY_SIZE(lg->idt))
 354                 set_trap(lg, &lg->idt[num], num, lo, hi);
 355         else if (num == SYSCALL_VECTOR)
 356                 set_trap(lg, &lg->syscall_idt, num, lo, hi);
 357 }
 358
 359 /* The default entry for each interrupt points into the Switcher routines which
 360  * simply return to the Host.  The run_guest() loop will then call
 361  * deliver_trap() to bounce it back into the Guest. */
 362 static void default_idt_entry(struct desc_struct *idt,
 363                               int trap,
 364                               const unsigned long handler)
 365 {
 366         /* A present interrupt gate. */
 367         u32 flags = 0x8e00;
 368
 369         /* Set the privilege level on the entry for the hypercall: this allows
 370          * the Guest to use the "int" instruction to trigger it. */
 371         if (trap == LGUEST_TRAP_ENTRY)
 372                 flags |= (GUEST_PL << 13);
 373
 374         /* Now pack it into the IDT entry in its weird format. */
 375         idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
 376         idt->b = (handler&0xFFFF0000) | flags;
 377 }
 378
 379 /* When the Guest first starts, we put default entries into the IDT. */
 380 void setup_default_idt_entries(struct lguest_ro_state *state,
 381                                const unsigned long *def)
 382 {
 383         unsigned int i;
 384
 385         for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
 386                 default_idt_entry(&state->guest_idt[i], i, def[i]);
 387 }
 388
 389 /*H:240 We don't use the IDT entries in the "struct lguest" directly, instead
 390  * we copy them into the IDT which we've set up for Guests on this CPU, just
 391  * before we run the Guest.  This routine does that copy. */
 392 void copy_traps(const struct lguest *lg, struct desc_struct *idt,
 393                 const unsigned long *def)
 394 {
 395         unsigned int i;
 396
 397         /* We can simply copy the direct traps, otherwise we use the default
 398          * ones in the Switcher: they will return to the Host. */
 399         for (i = 0; i < FIRST_EXTERNAL_VECTOR; i++) {
 400                 if (direct_trap(lg, &lg->idt[i], i))
 401                         idt[i] = lg->idt[i];
 402                 else
 403                         default_idt_entry(&idt[i], i, def[i]);
 404         }
 405
 406         /* Don't forget the system call trap!  The IDT entries for other
 407          * interupts never change, so no need to copy them. */
 408         i = SYSCALL_VECTOR;
 409         if (direct_trap(lg, &lg->syscall_idt, i))
 410                 idt[i] = lg->syscall_idt;
 411         else
 412                 default_idt_entry(&idt[i], i, def[i]);
 413 }
 414
 415 void guest_set_clockevent(struct lguest *lg, unsigned long delta)
 416 {
 417         ktime_t expires;
 418
 419         if (unlikely(delta == 0)) {
 420                 /* Clock event device is shutting down. */
 421                 hrtimer_cancel(&lg->hrt);
 422                 return;
 423         }
 424
 425         expires = ktime_add_ns(ktime_get_real(), delta);
 426         hrtimer_start(&lg->hrt, expires, HRTIMER_MODE_ABS);
 427 }
 428
 429 static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
 430 {
 431         struct lguest *lg = container_of(timer, struct lguest, hrt);
 432
 433         set_bit(0, lg->irqs_pending);
 434         if (lg->halted)
 435                 wake_up_process(lg->tsk);
 436         return HRTIMER_NORESTART;
 437 }
 438
 439 void init_clockdev(struct lguest *lg)
 440 {
 441         hrtimer_init(&lg->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
 442         lg->hrt.function = clockdev_fn;
 443 }