1 /*P:800 Interrupts (traps) are complicated enough to earn their own file.
2 * There are three classes of interrupts:
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.
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>
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);
23 /* The address of the interrupt handler is split into two bits: */
24 static unsigned long idt_address(u32 lo, u32 hi)
26 return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
29 /* The "type" of the interrupt handler is a 4 bit field: we only support a
31 static int idt_type(u32 lo, u32 hi)
33 return (hi >> 8) & 0xF;
36 /* An IDT entry can't be used unless the "present" bit is set. */
37 static bool idt_present(u32 lo, u32 hi)
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)
46 /* Stack grows upwards: move stack then write value. */
48 lgwrite(cpu, *gstack, u32, val);
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.
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.
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
63 static void set_guest_interrupt(struct lg_cpu *cpu, u32 lo, u32 hi,
66 unsigned long gstack, origstack;
67 u32 eflags, ss, irq_enable;
68 unsigned long virtstack;
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;
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);
87 /* We're staying on the same Guest (kernel) stack. */
88 virtstack = cpu->regs->esp;
91 origstack = gstack = guest_pa(cpu, virtstack);
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;
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
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);
110 /* For the six traps which supply an error code, we push that, too. */
112 push_guest_stack(cpu, &gstack, cpu->regs->errcode);
114 /* Now we've pushed all the old state, we change the stack, the code
115 * segment and the address to execute. */
117 cpu->regs->esp = virtstack + (gstack - origstack);
118 cpu->regs->cs = (__KERNEL_CS|GUEST_PL);
119 cpu->regs->eip = idt_address(lo, hi);
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");
129 * Virtual Interrupts.
131 * interrupt_pending() returns the first pending interrupt which isn't blocked
132 * by the Guest. It is called before every entry to the Guest, and just before
133 * we go to sleep when the Guest has halted itself. */
134 unsigned int interrupt_pending(struct lg_cpu *cpu, bool *more)
137 DECLARE_BITMAP(blk, LGUEST_IRQS);
139 /* If the Guest hasn't even initialized yet, we can do nothing. */
140 if (!cpu->lg->lguest_data)
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,
148 bitmap_andnot(blk, cpu->irqs_pending, blk, LGUEST_IRQS);
150 /* Find the first interrupt. */
151 irq = find_first_bit(blk, LGUEST_IRQS);
152 *more = find_next_bit(blk, LGUEST_IRQS, irq+1);
157 /* This actually diverts the Guest to running an interrupt handler, once an
158 * interrupt has been identified by interrupt_pending(). */
159 void try_deliver_interrupt(struct lg_cpu *cpu, unsigned int irq, bool more)
161 struct desc_struct *idt;
163 BUG_ON(irq >= LGUEST_IRQS);
165 /* They may be in the middle of an iret, where they asked us never to
166 * deliver interrupts. */
167 if (cpu->regs->eip >= cpu->lg->noirq_start &&
168 (cpu->regs->eip < cpu->lg->noirq_end))
171 /* If they're halted, interrupts restart them. */
173 /* Re-enable interrupts. */
174 if (put_user(X86_EFLAGS_IF, &cpu->lg->lguest_data->irq_enabled))
175 kill_guest(cpu, "Re-enabling interrupts");
178 /* Otherwise we check if they have interrupts disabled. */
180 if (get_user(irq_enabled, &cpu->lg->lguest_data->irq_enabled))
183 /* Make sure they know an IRQ is pending. */
184 put_user(X86_EFLAGS_IF,
185 &cpu->lg->lguest_data->irq_pending);
190 /* Look at the IDT entry the Guest gave us for this interrupt. The
191 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
193 idt = &cpu->arch.idt[FIRST_EXTERNAL_VECTOR+irq];
194 /* If they don't have a handler (yet?), we just ignore it */
195 if (idt_present(idt->a, idt->b)) {
196 /* OK, mark it no longer pending and deliver it. */
197 clear_bit(irq, cpu->irqs_pending);
198 /* set_guest_interrupt() takes the interrupt descriptor and a
199 * flag to say whether this interrupt pushes an error code onto
200 * the stack as well: virtual interrupts never do. */
201 set_guest_interrupt(cpu, idt->a, idt->b, false);
204 /* Every time we deliver an interrupt, we update the timestamp in the
205 * Guest's lguest_data struct. It would be better for the Guest if we
206 * did this more often, but it can actually be quite slow: doing it
207 * here is a compromise which means at least it gets updated every
208 * timer interrupt. */
209 write_timestamp(cpu);
211 /* If there are no other interrupts we want to deliver, clear
212 * the pending flag. */
214 put_user(0, &cpu->lg->lguest_data->irq_pending);
217 /* And this is the routine when we want to set an interrupt for the Guest. */
218 void set_interrupt(struct lg_cpu *cpu, unsigned int irq)
220 /* Next time the Guest runs, the core code will see if it can deliver
222 set_bit(irq, cpu->irqs_pending);
224 /* Make sure it sees it; it might be asleep (eg. halted), or
225 * running the Guest right now, in which case kick_process()
226 * will knock it out. */
227 if (!wake_up_process(cpu->tsk))
228 kick_process(cpu->tsk);
232 /* Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent
233 * me a patch, so we support that too. It'd be a big step for lguest if half
234 * the Plan 9 user base were to start using it.
236 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
237 * userbase. Oh well. */
238 static bool could_be_syscall(unsigned int num)
240 /* Normal Linux SYSCALL_VECTOR or reserved vector? */
241 return num == SYSCALL_VECTOR || num == syscall_vector;
244 /* The syscall vector it wants must be unused by Host. */
245 bool check_syscall_vector(struct lguest *lg)
249 if (get_user(vector, &lg->lguest_data->syscall_vec))
252 return could_be_syscall(vector);
255 int init_interrupts(void)
257 /* If they want some strange system call vector, reserve it now */
258 if (syscall_vector != SYSCALL_VECTOR) {
259 if (test_bit(syscall_vector, used_vectors) ||
260 vector_used_by_percpu_irq(syscall_vector)) {
261 printk(KERN_ERR "lg: couldn't reserve syscall %u\n",
265 set_bit(syscall_vector, used_vectors);
271 void free_interrupts(void)
273 if (syscall_vector != SYSCALL_VECTOR)
274 clear_bit(syscall_vector, used_vectors);
277 /*H:220 Now we've got the routines to deliver interrupts, delivering traps like
278 * page fault is easy. The only trick is that Intel decided that some traps
279 * should have error codes: */
280 static bool has_err(unsigned int trap)
282 return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
285 /* deliver_trap() returns true if it could deliver the trap. */
286 bool deliver_trap(struct lg_cpu *cpu, unsigned int num)
288 /* Trap numbers are always 8 bit, but we set an impossible trap number
289 * for traps inside the Switcher, so check that here. */
290 if (num >= ARRAY_SIZE(cpu->arch.idt))
293 /* Early on the Guest hasn't set the IDT entries (or maybe it put a
294 * bogus one in): if we fail here, the Guest will be killed. */
295 if (!idt_present(cpu->arch.idt[num].a, cpu->arch.idt[num].b))
297 set_guest_interrupt(cpu, cpu->arch.idt[num].a,
298 cpu->arch.idt[num].b, has_err(num));
302 /*H:250 Here's the hard part: returning to the Host every time a trap happens
303 * and then calling deliver_trap() and re-entering the Guest is slow.
304 * Particularly because Guest userspace system calls are traps (usually trap
307 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
308 * into the Guest. This is possible, but the complexities cause the size of
309 * this file to double! However, 150 lines of code is worth writing for taking
310 * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
311 * the other hypervisors would beat it up at lunchtime.
313 * This routine indicates if a particular trap number could be delivered
315 static bool direct_trap(unsigned int num)
317 /* Hardware interrupts don't go to the Guest at all (except system
319 if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
322 /* The Host needs to see page faults (for shadow paging and to save the
323 * fault address), general protection faults (in/out emulation) and
324 * device not available (TS handling), invalid opcode fault (kvm hcall),
325 * and of course, the hypercall trap. */
326 return num != 14 && num != 13 && num != 7 &&
327 num != 6 && num != LGUEST_TRAP_ENTRY;
331 /*M:005 The Guest has the ability to turn its interrupt gates into trap gates,
332 * if it is careful. The Host will let trap gates can go directly to the
333 * Guest, but the Guest needs the interrupts atomically disabled for an
334 * interrupt gate. It can do this by pointing the trap gate at instructions
335 * within noirq_start and noirq_end, where it can safely disable interrupts. */
337 /*M:006 The Guests do not use the sysenter (fast system call) instruction,
338 * because it's hardcoded to enter privilege level 0 and so can't go direct.
339 * It's about twice as fast as the older "int 0x80" system call, so it might
340 * still be worthwhile to handle it in the Switcher and lcall down to the
341 * Guest. The sysenter semantics are hairy tho: search for that keyword in
344 /*H:260 When we make traps go directly into the Guest, we need to make sure
345 * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the
346 * CPU trying to deliver the trap will fault while trying to push the interrupt
347 * words on the stack: this is called a double fault, and it forces us to kill
350 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */
351 void pin_stack_pages(struct lg_cpu *cpu)
355 /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or
356 * two pages of stack space. */
357 for (i = 0; i < cpu->lg->stack_pages; i++)
358 /* The stack grows *upwards*, so the address we're given is the
359 * start of the page after the kernel stack. Subtract one to
360 * get back onto the first stack page, and keep subtracting to
361 * get to the rest of the stack pages. */
362 pin_page(cpu, cpu->esp1 - 1 - i * PAGE_SIZE);
365 /* Direct traps also mean that we need to know whenever the Guest wants to use
366 * a different kernel stack, so we can change the IDT entries to use that
367 * stack. The IDT entries expect a virtual address, so unlike most addresses
368 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
371 * In Linux each process has its own kernel stack, so this happens a lot: we
372 * change stacks on each context switch. */
373 void guest_set_stack(struct lg_cpu *cpu, u32 seg, u32 esp, unsigned int pages)
375 /* You are not allowed have a stack segment with privilege level 0: bad
377 if ((seg & 0x3) != GUEST_PL)
378 kill_guest(cpu, "bad stack segment %i", seg);
379 /* We only expect one or two stack pages. */
381 kill_guest(cpu, "bad stack pages %u", pages);
382 /* Save where the stack is, and how many pages */
385 cpu->lg->stack_pages = pages;
386 /* Make sure the new stack pages are mapped */
387 pin_stack_pages(cpu);
390 /* All this reference to mapping stacks leads us neatly into the other complex
391 * part of the Host: page table handling. */
393 /*H:235 This is the routine which actually checks the Guest's IDT entry and
394 * transfers it into the entry in "struct lguest": */
395 static void set_trap(struct lg_cpu *cpu, struct desc_struct *trap,
396 unsigned int num, u32 lo, u32 hi)
398 u8 type = idt_type(lo, hi);
400 /* We zero-out a not-present entry */
401 if (!idt_present(lo, hi)) {
402 trap->a = trap->b = 0;
406 /* We only support interrupt and trap gates. */
407 if (type != 0xE && type != 0xF)
408 kill_guest(cpu, "bad IDT type %i", type);
410 /* We only copy the handler address, present bit, privilege level and
411 * type. The privilege level controls where the trap can be triggered
412 * manually with an "int" instruction. This is usually GUEST_PL,
413 * except for system calls which userspace can use. */
414 trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
415 trap->b = (hi&0xFFFFEF00);
418 /*H:230 While we're here, dealing with delivering traps and interrupts to the
419 * Guest, we might as well complete the picture: how the Guest tells us where
420 * it wants them to go. This would be simple, except making traps fast
421 * requires some tricks.
423 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
424 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */
425 void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int num, u32 lo, u32 hi)
427 /* Guest never handles: NMI, doublefault, spurious interrupt or
428 * hypercall. We ignore when it tries to set them. */
429 if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
432 /* Mark the IDT as changed: next time the Guest runs we'll know we have
433 * to copy this again. */
434 cpu->changed |= CHANGED_IDT;
436 /* Check that the Guest doesn't try to step outside the bounds. */
437 if (num >= ARRAY_SIZE(cpu->arch.idt))
438 kill_guest(cpu, "Setting idt entry %u", num);
440 set_trap(cpu, &cpu->arch.idt[num], num, lo, hi);
443 /* The default entry for each interrupt points into the Switcher routines which
444 * simply return to the Host. The run_guest() loop will then call
445 * deliver_trap() to bounce it back into the Guest. */
446 static void default_idt_entry(struct desc_struct *idt,
448 const unsigned long handler,
449 const struct desc_struct *base)
451 /* A present interrupt gate. */
454 /* Set the privilege level on the entry for the hypercall: this allows
455 * the Guest to use the "int" instruction to trigger it. */
456 if (trap == LGUEST_TRAP_ENTRY)
457 flags |= (GUEST_PL << 13);
459 /* Copy priv. level from what Guest asked for. This allows
460 * debug (int 3) traps from Guest userspace, for example. */
461 flags |= (base->b & 0x6000);
463 /* Now pack it into the IDT entry in its weird format. */
464 idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
465 idt->b = (handler&0xFFFF0000) | flags;
468 /* When the Guest first starts, we put default entries into the IDT. */
469 void setup_default_idt_entries(struct lguest_ro_state *state,
470 const unsigned long *def)
474 for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
475 default_idt_entry(&state->guest_idt[i], i, def[i], NULL);
478 /*H:240 We don't use the IDT entries in the "struct lguest" directly, instead
479 * we copy them into the IDT which we've set up for Guests on this CPU, just
480 * before we run the Guest. This routine does that copy. */
481 void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt,
482 const unsigned long *def)
486 /* We can simply copy the direct traps, otherwise we use the default
487 * ones in the Switcher: they will return to the Host. */
488 for (i = 0; i < ARRAY_SIZE(cpu->arch.idt); i++) {
489 const struct desc_struct *gidt = &cpu->arch.idt[i];
491 /* If no Guest can ever override this trap, leave it alone. */
495 /* Only trap gates (type 15) can go direct to the Guest.
496 * Interrupt gates (type 14) disable interrupts as they are
497 * entered, which we never let the Guest do. Not present
498 * entries (type 0x0) also can't go direct, of course.
500 * If it can't go direct, we still need to copy the priv. level:
501 * they might want to give userspace access to a software
503 if (idt_type(gidt->a, gidt->b) == 0xF)
506 default_idt_entry(&idt[i], i, def[i], gidt);
513 * There are two sources of virtual interrupts. We saw one in lguest_user.c:
514 * the Launcher sending interrupts for virtual devices. The other is the Guest
517 * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
518 * the next timer interrupt (in nanoseconds). We use the high-resolution timer
519 * infrastructure to set a callback at that time.
521 * 0 means "turn off the clock". */
522 void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta)
526 if (unlikely(delta == 0)) {
527 /* Clock event device is shutting down. */
528 hrtimer_cancel(&cpu->hrt);
532 /* We use wallclock time here, so the Guest might not be running for
533 * all the time between now and the timer interrupt it asked for. This
534 * is almost always the right thing to do. */
535 expires = ktime_add_ns(ktime_get_real(), delta);
536 hrtimer_start(&cpu->hrt, expires, HRTIMER_MODE_ABS);
539 /* This is the function called when the Guest's timer expires. */
540 static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
542 struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt);
544 /* Remember the first interrupt is the timer interrupt. */
545 set_interrupt(cpu, 0);
546 return HRTIMER_NORESTART;
549 /* This sets up the timer for this Guest. */
550 void init_clockdev(struct lg_cpu *cpu)
552 hrtimer_init(&cpu->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
553 cpu->hrt.function = clockdev_fn;