2 * Interrupts (traps) are complicated enough to earn their own file.
3 * There are three classes of interrupts:
5 * 1) Real hardware interrupts which occur while we're running the Guest,
6 * 2) Interrupts for virtual devices attached to the Guest, and
7 * 3) Traps and faults from the Guest.
9 * Real hardware interrupts must be delivered to the Host, not the Guest.
10 * Virtual interrupts must be delivered to the Guest, but we make them look
11 * just like real hardware would deliver them. Traps from the Guest can be set
12 * up to go directly back into the Guest, but sometimes the Host wants to see
13 * them first, so we also have a way of "reflecting" them into the Guest as if
14 * they had been delivered to it directly.
16 #include <linux/uaccess.h>
17 #include <linux/interrupt.h>
18 #include <linux/module.h>
21 /* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */
22 static unsigned int syscall_vector
= SYSCALL_VECTOR
;
23 module_param(syscall_vector
, uint
, 0444);
25 /* The address of the interrupt handler is split into two bits: */
26 static unsigned long idt_address(u32 lo
, u32 hi
)
28 return (lo
& 0x0000FFFF) | (hi
& 0xFFFF0000);
32 * The "type" of the interrupt handler is a 4 bit field: we only support a
35 static int idt_type(u32 lo
, u32 hi
)
37 return (hi
>> 8) & 0xF;
40 /* An IDT entry can't be used unless the "present" bit is set. */
41 static bool idt_present(u32 lo
, u32 hi
)
47 * We need a helper to "push" a value onto the Guest's stack, since that's a
48 * big part of what delivering an interrupt does.
50 static void push_guest_stack(struct lg_cpu
*cpu
, unsigned long *gstack
, u32 val
)
52 /* Stack grows upwards: move stack then write value. */
54 lgwrite(cpu
, *gstack
, u32
, val
);
58 * The set_guest_interrupt() routine actually delivers the interrupt or
59 * trap. The mechanics of delivering traps and interrupts to the Guest are the
60 * same, except some traps have an "error code" which gets pushed onto the
61 * stack as well: the caller tells us if this is one.
63 * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
64 * interrupt or trap. It's split into two parts for traditional reasons: gcc
65 * on i386 used to be frightened by 64 bit numbers.
67 * We set up the stack just like the CPU does for a real interrupt, so it's
68 * identical for the Guest (and the standard "iret" instruction will undo
71 static void set_guest_interrupt(struct lg_cpu
*cpu
, u32 lo
, u32 hi
,
74 unsigned long gstack
, origstack
;
75 u32 eflags
, ss
, irq_enable
;
76 unsigned long virtstack
;
79 * There are two cases for interrupts: one where the Guest is already
80 * in the kernel, and a more complex one where the Guest is in
81 * userspace. We check the privilege level to find out.
83 if ((cpu
->regs
->ss
&0x3) != GUEST_PL
) {
85 * The Guest told us their kernel stack with the SET_STACK
86 * hypercall: both the virtual address and the segment.
88 virtstack
= cpu
->esp1
;
91 origstack
= gstack
= guest_pa(cpu
, virtstack
);
93 * We push the old stack segment and pointer onto the new
94 * stack: when the Guest does an "iret" back from the interrupt
95 * handler the CPU will notice they're dropping privilege
96 * levels and expect these here.
98 push_guest_stack(cpu
, &gstack
, cpu
->regs
->ss
);
99 push_guest_stack(cpu
, &gstack
, cpu
->regs
->esp
);
101 /* We're staying on the same Guest (kernel) stack. */
102 virtstack
= cpu
->regs
->esp
;
105 origstack
= gstack
= guest_pa(cpu
, virtstack
);
109 * Remember that we never let the Guest actually disable interrupts, so
110 * the "Interrupt Flag" bit is always set. We copy that bit from the
111 * Guest's "irq_enabled" field into the eflags word: we saw the Guest
112 * copy it back in "lguest_iret".
114 eflags
= cpu
->regs
->eflags
;
115 if (get_user(irq_enable
, &cpu
->lg
->lguest_data
->irq_enabled
) == 0
116 && !(irq_enable
& X86_EFLAGS_IF
))
117 eflags
&= ~X86_EFLAGS_IF
;
120 * An interrupt is expected to push three things on the stack: the old
121 * "eflags" word, the old code segment, and the old instruction
124 push_guest_stack(cpu
, &gstack
, eflags
);
125 push_guest_stack(cpu
, &gstack
, cpu
->regs
->cs
);
126 push_guest_stack(cpu
, &gstack
, cpu
->regs
->eip
);
128 /* For the six traps which supply an error code, we push that, too. */
130 push_guest_stack(cpu
, &gstack
, cpu
->regs
->errcode
);
133 * Now we've pushed all the old state, we change the stack, the code
134 * segment and the address to execute.
137 cpu
->regs
->esp
= virtstack
+ (gstack
- origstack
);
138 cpu
->regs
->cs
= (__KERNEL_CS
|GUEST_PL
);
139 cpu
->regs
->eip
= idt_address(lo
, hi
);
142 * There are two kinds of interrupt handlers: 0xE is an "interrupt
143 * gate" which expects interrupts to be disabled on entry.
145 if (idt_type(lo
, hi
) == 0xE)
146 if (put_user(0, &cpu
->lg
->lguest_data
->irq_enabled
))
147 kill_guest(cpu
, "Disabling interrupts");
151 * Virtual Interrupts.
153 * interrupt_pending() returns the first pending interrupt which isn't blocked
154 * by the Guest. It is called before every entry to the Guest, and just before
155 * we go to sleep when the Guest has halted itself.
157 unsigned int interrupt_pending(struct lg_cpu
*cpu
, bool *more
)
160 DECLARE_BITMAP(blk
, LGUEST_IRQS
);
162 /* If the Guest hasn't even initialized yet, we can do nothing. */
163 if (!cpu
->lg
->lguest_data
)
167 * Take our "irqs_pending" array and remove any interrupts the Guest
168 * wants blocked: the result ends up in "blk".
170 if (copy_from_user(&blk
, cpu
->lg
->lguest_data
->blocked_interrupts
,
173 bitmap_andnot(blk
, cpu
->irqs_pending
, blk
, LGUEST_IRQS
);
175 /* Find the first interrupt. */
176 irq
= find_first_bit(blk
, LGUEST_IRQS
);
177 *more
= find_next_bit(blk
, LGUEST_IRQS
, irq
+1);
183 * This actually diverts the Guest to running an interrupt handler, once an
184 * interrupt has been identified by interrupt_pending().
186 void try_deliver_interrupt(struct lg_cpu
*cpu
, unsigned int irq
, bool more
)
188 struct desc_struct
*idt
;
190 BUG_ON(irq
>= LGUEST_IRQS
);
193 * They may be in the middle of an iret, where they asked us never to
194 * deliver interrupts.
196 if (cpu
->regs
->eip
>= cpu
->lg
->noirq_start
&&
197 (cpu
->regs
->eip
< cpu
->lg
->noirq_end
))
200 /* If they're halted, interrupts restart them. */
202 /* Re-enable interrupts. */
203 if (put_user(X86_EFLAGS_IF
, &cpu
->lg
->lguest_data
->irq_enabled
))
204 kill_guest(cpu
, "Re-enabling interrupts");
207 /* Otherwise we check if they have interrupts disabled. */
209 if (get_user(irq_enabled
, &cpu
->lg
->lguest_data
->irq_enabled
))
212 /* Make sure they know an IRQ is pending. */
213 put_user(X86_EFLAGS_IF
,
214 &cpu
->lg
->lguest_data
->irq_pending
);
220 * Look at the IDT entry the Guest gave us for this interrupt. The
221 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
224 idt
= &cpu
->arch
.idt
[FIRST_EXTERNAL_VECTOR
+irq
];
225 /* If they don't have a handler (yet?), we just ignore it */
226 if (idt_present(idt
->a
, idt
->b
)) {
227 /* OK, mark it no longer pending and deliver it. */
228 clear_bit(irq
, cpu
->irqs_pending
);
230 * set_guest_interrupt() takes the interrupt descriptor and a
231 * flag to say whether this interrupt pushes an error code onto
232 * the stack as well: virtual interrupts never do.
234 set_guest_interrupt(cpu
, idt
->a
, idt
->b
, false);
238 * Every time we deliver an interrupt, we update the timestamp in the
239 * Guest's lguest_data struct. It would be better for the Guest if we
240 * did this more often, but it can actually be quite slow: doing it
241 * here is a compromise which means at least it gets updated every
244 write_timestamp(cpu
);
247 * If there are no other interrupts we want to deliver, clear
251 put_user(0, &cpu
->lg
->lguest_data
->irq_pending
);
254 /* And this is the routine when we want to set an interrupt for the Guest. */
255 void set_interrupt(struct lg_cpu
*cpu
, unsigned int irq
)
258 * Next time the Guest runs, the core code will see if it can deliver
261 set_bit(irq
, cpu
->irqs_pending
);
264 * Make sure it sees it; it might be asleep (eg. halted), or running
265 * the Guest right now, in which case kick_process() will knock it out.
267 if (!wake_up_process(cpu
->tsk
))
268 kick_process(cpu
->tsk
);
273 * Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent
274 * me a patch, so we support that too. It'd be a big step for lguest if half
275 * the Plan 9 user base were to start using it.
277 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
280 static bool could_be_syscall(unsigned int num
)
282 /* Normal Linux SYSCALL_VECTOR or reserved vector? */
283 return num
== SYSCALL_VECTOR
|| num
== syscall_vector
;
286 /* The syscall vector it wants must be unused by Host. */
287 bool check_syscall_vector(struct lguest
*lg
)
291 if (get_user(vector
, &lg
->lguest_data
->syscall_vec
))
294 return could_be_syscall(vector
);
297 int init_interrupts(void)
299 /* If they want some strange system call vector, reserve it now */
300 if (syscall_vector
!= SYSCALL_VECTOR
) {
301 if (test_bit(syscall_vector
, used_vectors
) ||
302 vector_used_by_percpu_irq(syscall_vector
)) {
303 printk(KERN_ERR
"lg: couldn't reserve syscall %u\n",
307 set_bit(syscall_vector
, used_vectors
);
313 void free_interrupts(void)
315 if (syscall_vector
!= SYSCALL_VECTOR
)
316 clear_bit(syscall_vector
, used_vectors
);
320 * Now we've got the routines to deliver interrupts, delivering traps like
321 * page fault is easy. The only trick is that Intel decided that some traps
322 * should have error codes:
324 static bool has_err(unsigned int trap
)
326 return (trap
== 8 || (trap
>= 10 && trap
<= 14) || trap
== 17);
329 /* deliver_trap() returns true if it could deliver the trap. */
330 bool deliver_trap(struct lg_cpu
*cpu
, unsigned int num
)
333 * Trap numbers are always 8 bit, but we set an impossible trap number
334 * for traps inside the Switcher, so check that here.
336 if (num
>= ARRAY_SIZE(cpu
->arch
.idt
))
340 * Early on the Guest hasn't set the IDT entries (or maybe it put a
341 * bogus one in): if we fail here, the Guest will be killed.
343 if (!idt_present(cpu
->arch
.idt
[num
].a
, cpu
->arch
.idt
[num
].b
))
345 set_guest_interrupt(cpu
, cpu
->arch
.idt
[num
].a
,
346 cpu
->arch
.idt
[num
].b
, has_err(num
));
351 * Here's the hard part: returning to the Host every time a trap happens
352 * and then calling deliver_trap() and re-entering the Guest is slow.
353 * Particularly because Guest userspace system calls are traps (usually trap
356 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
357 * into the Guest. This is possible, but the complexities cause the size of
358 * this file to double! However, 150 lines of code is worth writing for taking
359 * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
360 * the other hypervisors would beat it up at lunchtime.
362 * This routine indicates if a particular trap number could be delivered
365 static bool direct_trap(unsigned int num
)
368 * Hardware interrupts don't go to the Guest at all (except system
371 if (num
>= FIRST_EXTERNAL_VECTOR
&& !could_be_syscall(num
))
375 * The Host needs to see page faults (for shadow paging and to save the
376 * fault address), general protection faults (in/out emulation) and
377 * device not available (TS handling), invalid opcode fault (kvm hcall),
378 * and of course, the hypercall trap.
380 return num
!= 14 && num
!= 13 && num
!= 7 &&
381 num
!= 6 && num
!= LGUEST_TRAP_ENTRY
;
386 * The Guest has the ability to turn its interrupt gates into trap gates,
387 * if it is careful. The Host will let trap gates can go directly to the
388 * Guest, but the Guest needs the interrupts atomically disabled for an
389 * interrupt gate. It can do this by pointing the trap gate at instructions
390 * within noirq_start and noirq_end, where it can safely disable interrupts.
394 * The Guests do not use the sysenter (fast system call) instruction,
395 * because it's hardcoded to enter privilege level 0 and so can't go direct.
396 * It's about twice as fast as the older "int 0x80" system call, so it might
397 * still be worthwhile to handle it in the Switcher and lcall down to the
398 * Guest. The sysenter semantics are hairy tho: search for that keyword in
403 * When we make traps go directly into the Guest, we need to make sure
404 * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the
405 * CPU trying to deliver the trap will fault while trying to push the interrupt
406 * words on the stack: this is called a double fault, and it forces us to kill
409 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault.
411 void pin_stack_pages(struct lg_cpu
*cpu
)
416 * Depending on the CONFIG_4KSTACKS option, the Guest can have one or
417 * two pages of stack space.
419 for (i
= 0; i
< cpu
->lg
->stack_pages
; i
++)
421 * The stack grows *upwards*, so the address we're given is the
422 * start of the page after the kernel stack. Subtract one to
423 * get back onto the first stack page, and keep subtracting to
424 * get to the rest of the stack pages.
426 pin_page(cpu
, cpu
->esp1
- 1 - i
* PAGE_SIZE
);
430 * Direct traps also mean that we need to know whenever the Guest wants to use
431 * a different kernel stack, so we can change the IDT entries to use that
432 * stack. The IDT entries expect a virtual address, so unlike most addresses
433 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
436 * In Linux each process has its own kernel stack, so this happens a lot: we
437 * change stacks on each context switch.
439 void guest_set_stack(struct lg_cpu
*cpu
, u32 seg
, u32 esp
, unsigned int pages
)
442 * You're not allowed a stack segment with privilege level 0: bad Guest!
444 if ((seg
& 0x3) != GUEST_PL
)
445 kill_guest(cpu
, "bad stack segment %i", seg
);
446 /* We only expect one or two stack pages. */
448 kill_guest(cpu
, "bad stack pages %u", pages
);
449 /* Save where the stack is, and how many pages */
452 cpu
->lg
->stack_pages
= pages
;
453 /* Make sure the new stack pages are mapped */
454 pin_stack_pages(cpu
);
458 * All this reference to mapping stacks leads us neatly into the other complex
459 * part of the Host: page table handling.
463 * This is the routine which actually checks the Guest's IDT entry and
464 * transfers it into the entry in "struct lguest":
466 static void set_trap(struct lg_cpu
*cpu
, struct desc_struct
*trap
,
467 unsigned int num
, u32 lo
, u32 hi
)
469 u8 type
= idt_type(lo
, hi
);
471 /* We zero-out a not-present entry */
472 if (!idt_present(lo
, hi
)) {
473 trap
->a
= trap
->b
= 0;
477 /* We only support interrupt and trap gates. */
478 if (type
!= 0xE && type
!= 0xF)
479 kill_guest(cpu
, "bad IDT type %i", type
);
482 * We only copy the handler address, present bit, privilege level and
483 * type. The privilege level controls where the trap can be triggered
484 * manually with an "int" instruction. This is usually GUEST_PL,
485 * except for system calls which userspace can use.
487 trap
->a
= ((__KERNEL_CS
|GUEST_PL
)<<16) | (lo
&0x0000FFFF);
488 trap
->b
= (hi
&0xFFFFEF00);
492 * While we're here, dealing with delivering traps and interrupts to the
493 * Guest, we might as well complete the picture: how the Guest tells us where
494 * it wants them to go. This would be simple, except making traps fast
495 * requires some tricks.
497 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
498 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here.
500 void load_guest_idt_entry(struct lg_cpu
*cpu
, unsigned int num
, u32 lo
, u32 hi
)
503 * Guest never handles: NMI, doublefault, spurious interrupt or
504 * hypercall. We ignore when it tries to set them.
506 if (num
== 2 || num
== 8 || num
== 15 || num
== LGUEST_TRAP_ENTRY
)
510 * Mark the IDT as changed: next time the Guest runs we'll know we have
511 * to copy this again.
513 cpu
->changed
|= CHANGED_IDT
;
515 /* Check that the Guest doesn't try to step outside the bounds. */
516 if (num
>= ARRAY_SIZE(cpu
->arch
.idt
))
517 kill_guest(cpu
, "Setting idt entry %u", num
);
519 set_trap(cpu
, &cpu
->arch
.idt
[num
], num
, lo
, hi
);
523 * The default entry for each interrupt points into the Switcher routines which
524 * simply return to the Host. The run_guest() loop will then call
525 * deliver_trap() to bounce it back into the Guest.
527 static void default_idt_entry(struct desc_struct
*idt
,
529 const unsigned long handler
,
530 const struct desc_struct
*base
)
532 /* A present interrupt gate. */
536 * Set the privilege level on the entry for the hypercall: this allows
537 * the Guest to use the "int" instruction to trigger it.
539 if (trap
== LGUEST_TRAP_ENTRY
)
540 flags
|= (GUEST_PL
<< 13);
543 * Copy privilege level from what Guest asked for. This allows
544 * debug (int 3) traps from Guest userspace, for example.
546 flags
|= (base
->b
& 0x6000);
548 /* Now pack it into the IDT entry in its weird format. */
549 idt
->a
= (LGUEST_CS
<<16) | (handler
&0x0000FFFF);
550 idt
->b
= (handler
&0xFFFF0000) | flags
;
553 /* When the Guest first starts, we put default entries into the IDT. */
554 void setup_default_idt_entries(struct lguest_ro_state
*state
,
555 const unsigned long *def
)
559 for (i
= 0; i
< ARRAY_SIZE(state
->guest_idt
); i
++)
560 default_idt_entry(&state
->guest_idt
[i
], i
, def
[i
], NULL
);
564 * We don't use the IDT entries in the "struct lguest" directly, instead
565 * we copy them into the IDT which we've set up for Guests on this CPU, just
566 * before we run the Guest. This routine does that copy.
568 void copy_traps(const struct lg_cpu
*cpu
, struct desc_struct
*idt
,
569 const unsigned long *def
)
574 * We can simply copy the direct traps, otherwise we use the default
575 * ones in the Switcher: they will return to the Host.
577 for (i
= 0; i
< ARRAY_SIZE(cpu
->arch
.idt
); i
++) {
578 const struct desc_struct
*gidt
= &cpu
->arch
.idt
[i
];
580 /* If no Guest can ever override this trap, leave it alone. */
585 * Only trap gates (type 15) can go direct to the Guest.
586 * Interrupt gates (type 14) disable interrupts as they are
587 * entered, which we never let the Guest do. Not present
588 * entries (type 0x0) also can't go direct, of course.
590 * If it can't go direct, we still need to copy the priv. level:
591 * they might want to give userspace access to a software
594 if (idt_type(gidt
->a
, gidt
->b
) == 0xF)
597 default_idt_entry(&idt
[i
], i
, def
[i
], gidt
);
604 * There are two sources of virtual interrupts. We saw one in lguest_user.c:
605 * the Launcher sending interrupts for virtual devices. The other is the Guest
608 * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
609 * the next timer interrupt (in nanoseconds). We use the high-resolution timer
610 * infrastructure to set a callback at that time.
612 * 0 means "turn off the clock".
614 void guest_set_clockevent(struct lg_cpu
*cpu
, unsigned long delta
)
618 if (unlikely(delta
== 0)) {
619 /* Clock event device is shutting down. */
620 hrtimer_cancel(&cpu
->hrt
);
625 * We use wallclock time here, so the Guest might not be running for
626 * all the time between now and the timer interrupt it asked for. This
627 * is almost always the right thing to do.
629 expires
= ktime_add_ns(ktime_get_real(), delta
);
630 hrtimer_start(&cpu
->hrt
, expires
, HRTIMER_MODE_ABS
);
633 /* This is the function called when the Guest's timer expires. */
634 static enum hrtimer_restart
clockdev_fn(struct hrtimer
*timer
)
636 struct lg_cpu
*cpu
= container_of(timer
, struct lg_cpu
, hrt
);
638 /* Remember the first interrupt is the timer interrupt. */
639 set_interrupt(cpu
, 0);
640 return HRTIMER_NORESTART
;
643 /* This sets up the timer for this Guest. */
644 void init_clockdev(struct lg_cpu
*cpu
)
646 hrtimer_init(&cpu
->hrt
, CLOCK_REALTIME
, HRTIMER_MODE_ABS
);
647 cpu
->hrt
.function
= clockdev_fn
;
This page took 0.044295 seconds and 5 git commands to generate.