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>
19 #include <linux/sched.h>
22 /* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */
23 static unsigned int syscall_vector
= IA32_SYSCALL_VECTOR
;
24 module_param(syscall_vector
, uint
, 0444);
26 /* The address of the interrupt handler is split into two bits: */
27 static unsigned long idt_address(u32 lo
, u32 hi
)
29 return (lo
& 0x0000FFFF) | (hi
& 0xFFFF0000);
33 * The "type" of the interrupt handler is a 4 bit field: we only support a
36 static int idt_type(u32 lo
, u32 hi
)
38 return (hi
>> 8) & 0xF;
41 /* An IDT entry can't be used unless the "present" bit is set. */
42 static bool idt_present(u32 lo
, u32 hi
)
48 * We need a helper to "push" a value onto the Guest's stack, since that's a
49 * big part of what delivering an interrupt does.
51 static void push_guest_stack(struct lg_cpu
*cpu
, unsigned long *gstack
, u32 val
)
53 /* Stack grows upwards: move stack then write value. */
55 lgwrite(cpu
, *gstack
, u32
, val
);
59 * The push_guest_interrupt_stack() routine saves Guest state on the stack for
60 * an interrupt or trap. The mechanics of delivering traps and interrupts to
61 * the Guest are the same, except some traps have an "error code" which gets
62 * pushed onto the stack as well: the caller tells us if this is one.
64 * We set up the stack just like the CPU does for a real interrupt, so it's
65 * identical for the Guest (and the standard "iret" instruction will undo
68 static void push_guest_interrupt_stack(struct lg_cpu
*cpu
, bool has_err
)
70 unsigned long gstack
, origstack
;
71 u32 eflags
, ss
, irq_enable
;
72 unsigned long virtstack
;
75 * There are two cases for interrupts: one where the Guest is already
76 * in the kernel, and a more complex one where the Guest is in
77 * userspace. We check the privilege level to find out.
79 if ((cpu
->regs
->ss
&0x3) != GUEST_PL
) {
81 * The Guest told us their kernel stack with the SET_STACK
82 * hypercall: both the virtual address and the segment.
84 virtstack
= cpu
->esp1
;
87 origstack
= gstack
= guest_pa(cpu
, virtstack
);
89 * We push the old stack segment and pointer onto the new
90 * stack: when the Guest does an "iret" back from the interrupt
91 * handler the CPU will notice they're dropping privilege
92 * levels and expect these here.
94 push_guest_stack(cpu
, &gstack
, cpu
->regs
->ss
);
95 push_guest_stack(cpu
, &gstack
, cpu
->regs
->esp
);
97 /* We're staying on the same Guest (kernel) stack. */
98 virtstack
= cpu
->regs
->esp
;
101 origstack
= gstack
= guest_pa(cpu
, virtstack
);
105 * Remember that we never let the Guest actually disable interrupts, so
106 * the "Interrupt Flag" bit is always set. We copy that bit from the
107 * Guest's "irq_enabled" field into the eflags word: we saw the Guest
108 * copy it back in "lguest_iret".
110 eflags
= cpu
->regs
->eflags
;
111 if (get_user(irq_enable
, &cpu
->lg
->lguest_data
->irq_enabled
) == 0
112 && !(irq_enable
& X86_EFLAGS_IF
))
113 eflags
&= ~X86_EFLAGS_IF
;
116 * An interrupt is expected to push three things on the stack: the old
117 * "eflags" word, the old code segment, and the old instruction
120 push_guest_stack(cpu
, &gstack
, eflags
);
121 push_guest_stack(cpu
, &gstack
, cpu
->regs
->cs
);
122 push_guest_stack(cpu
, &gstack
, cpu
->regs
->eip
);
124 /* For the six traps which supply an error code, we push that, too. */
126 push_guest_stack(cpu
, &gstack
, cpu
->regs
->errcode
);
128 /* Adjust the stack pointer and stack segment. */
130 cpu
->regs
->esp
= virtstack
+ (gstack
- origstack
);
134 * This actually makes the Guest start executing the given interrupt/trap
137 * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
138 * interrupt or trap. It's split into two parts for traditional reasons: gcc
139 * on i386 used to be frightened by 64 bit numbers.
141 static void guest_run_interrupt(struct lg_cpu
*cpu
, u32 lo
, u32 hi
)
143 /* If we're already in the kernel, we don't change stacks. */
144 if ((cpu
->regs
->ss
&0x3) != GUEST_PL
)
145 cpu
->regs
->ss
= cpu
->esp1
;
148 * Set the code segment and the address to execute.
150 cpu
->regs
->cs
= (__KERNEL_CS
|GUEST_PL
);
151 cpu
->regs
->eip
= idt_address(lo
, hi
);
154 * Trapping always clears these flags:
156 * VM: Virtual 8086 mode
161 ~(X86_EFLAGS_TF
|X86_EFLAGS_VM
|X86_EFLAGS_RF
|X86_EFLAGS_NT
);
164 * There are two kinds of interrupt handlers: 0xE is an "interrupt
165 * gate" which expects interrupts to be disabled on entry.
167 if (idt_type(lo
, hi
) == 0xE)
168 if (put_user(0, &cpu
->lg
->lguest_data
->irq_enabled
))
169 kill_guest(cpu
, "Disabling interrupts");
172 /* This restores the eflags word which was pushed on the stack by a trap */
173 static void restore_eflags(struct lg_cpu
*cpu
)
175 /* This is the physical address of the stack. */
176 unsigned long stack_pa
= guest_pa(cpu
, cpu
->regs
->esp
);
179 * Stack looks like this:
185 cpu
->regs
->eflags
= lgread(cpu
, stack_pa
+ 8, u32
);
187 ~(X86_EFLAGS_TF
|X86_EFLAGS_VM
|X86_EFLAGS_RF
|X86_EFLAGS_NT
);
191 * Virtual Interrupts.
193 * interrupt_pending() returns the first pending interrupt which isn't blocked
194 * by the Guest. It is called before every entry to the Guest, and just before
195 * we go to sleep when the Guest has halted itself.
197 unsigned int interrupt_pending(struct lg_cpu
*cpu
, bool *more
)
200 DECLARE_BITMAP(blk
, LGUEST_IRQS
);
202 /* If the Guest hasn't even initialized yet, we can do nothing. */
203 if (!cpu
->lg
->lguest_data
)
207 * Take our "irqs_pending" array and remove any interrupts the Guest
208 * wants blocked: the result ends up in "blk".
210 if (copy_from_user(&blk
, cpu
->lg
->lguest_data
->blocked_interrupts
,
213 bitmap_andnot(blk
, cpu
->irqs_pending
, blk
, LGUEST_IRQS
);
215 /* Find the first interrupt. */
216 irq
= find_first_bit(blk
, LGUEST_IRQS
);
217 *more
= find_next_bit(blk
, LGUEST_IRQS
, irq
+1);
223 * This actually diverts the Guest to running an interrupt handler, once an
224 * interrupt has been identified by interrupt_pending().
226 void try_deliver_interrupt(struct lg_cpu
*cpu
, unsigned int irq
, bool more
)
228 struct desc_struct
*idt
;
230 BUG_ON(irq
>= LGUEST_IRQS
);
232 /* If they're halted, interrupts restart them. */
234 /* Re-enable interrupts. */
235 if (put_user(X86_EFLAGS_IF
, &cpu
->lg
->lguest_data
->irq_enabled
))
236 kill_guest(cpu
, "Re-enabling interrupts");
239 /* Otherwise we check if they have interrupts disabled. */
241 if (get_user(irq_enabled
, &cpu
->lg
->lguest_data
->irq_enabled
))
244 /* Make sure they know an IRQ is pending. */
245 put_user(X86_EFLAGS_IF
,
246 &cpu
->lg
->lguest_data
->irq_pending
);
252 * Look at the IDT entry the Guest gave us for this interrupt. The
253 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
256 idt
= &cpu
->arch
.idt
[FIRST_EXTERNAL_VECTOR
+irq
];
257 /* If they don't have a handler (yet?), we just ignore it */
258 if (idt_present(idt
->a
, idt
->b
)) {
259 /* OK, mark it no longer pending and deliver it. */
260 clear_bit(irq
, cpu
->irqs_pending
);
263 * They may be about to iret, where they asked us never to
264 * deliver interrupts. In this case, we can emulate that iret
265 * then immediately deliver the interrupt. This is basically
266 * a noop: the iret would pop the interrupt frame and restore
267 * eflags, and then we'd set it up again. So just restore the
268 * eflags word and jump straight to the handler in this case.
270 * Denys Vlasenko points out that this isn't quite right: if
271 * the iret was returning to userspace, then that interrupt
272 * would reset the stack pointer (which the Guest told us
273 * about via LHCALL_SET_STACK). But unless the Guest is being
274 * *really* weird, that will be the same as the current stack
277 if (cpu
->regs
->eip
== cpu
->lg
->noirq_iret
) {
281 * set_guest_interrupt() takes a flag to say whether
282 * this interrupt pushes an error code onto the stack
283 * as well: virtual interrupts never do.
285 push_guest_interrupt_stack(cpu
, false);
287 /* Actually make Guest cpu jump to handler. */
288 guest_run_interrupt(cpu
, idt
->a
, idt
->b
);
292 * Every time we deliver an interrupt, we update the timestamp in the
293 * Guest's lguest_data struct. It would be better for the Guest if we
294 * did this more often, but it can actually be quite slow: doing it
295 * here is a compromise which means at least it gets updated every
298 write_timestamp(cpu
);
301 * If there are no other interrupts we want to deliver, clear
305 put_user(0, &cpu
->lg
->lguest_data
->irq_pending
);
308 /* And this is the routine when we want to set an interrupt for the Guest. */
309 void set_interrupt(struct lg_cpu
*cpu
, unsigned int irq
)
312 * Next time the Guest runs, the core code will see if it can deliver
315 set_bit(irq
, cpu
->irqs_pending
);
318 * Make sure it sees it; it might be asleep (eg. halted), or running
319 * the Guest right now, in which case kick_process() will knock it out.
321 if (!wake_up_process(cpu
->tsk
))
322 kick_process(cpu
->tsk
);
327 * Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent
328 * me a patch, so we support that too. It'd be a big step for lguest if half
329 * the Plan 9 user base were to start using it.
331 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
334 bool could_be_syscall(unsigned int num
)
336 /* Normal Linux IA32_SYSCALL_VECTOR or reserved vector? */
337 return num
== IA32_SYSCALL_VECTOR
|| num
== syscall_vector
;
340 /* The syscall vector it wants must be unused by Host. */
341 bool check_syscall_vector(struct lguest
*lg
)
345 if (get_user(vector
, &lg
->lguest_data
->syscall_vec
))
348 return could_be_syscall(vector
);
351 int init_interrupts(void)
353 /* If they want some strange system call vector, reserve it now */
354 if (syscall_vector
!= IA32_SYSCALL_VECTOR
) {
355 if (test_bit(syscall_vector
, used_vectors
) ||
356 vector_used_by_percpu_irq(syscall_vector
)) {
357 printk(KERN_ERR
"lg: couldn't reserve syscall %u\n",
361 set_bit(syscall_vector
, used_vectors
);
367 void free_interrupts(void)
369 if (syscall_vector
!= IA32_SYSCALL_VECTOR
)
370 clear_bit(syscall_vector
, used_vectors
);
374 * Now we've got the routines to deliver interrupts, delivering traps like
375 * page fault is easy. The only trick is that Intel decided that some traps
376 * should have error codes:
378 static bool has_err(unsigned int trap
)
380 return (trap
== 8 || (trap
>= 10 && trap
<= 14) || trap
== 17);
383 /* deliver_trap() returns true if it could deliver the trap. */
384 bool deliver_trap(struct lg_cpu
*cpu
, unsigned int num
)
387 * Trap numbers are always 8 bit, but we set an impossible trap number
388 * for traps inside the Switcher, so check that here.
390 if (num
>= ARRAY_SIZE(cpu
->arch
.idt
))
394 * Early on the Guest hasn't set the IDT entries (or maybe it put a
395 * bogus one in): if we fail here, the Guest will be killed.
397 if (!idt_present(cpu
->arch
.idt
[num
].a
, cpu
->arch
.idt
[num
].b
))
399 push_guest_interrupt_stack(cpu
, has_err(num
));
400 guest_run_interrupt(cpu
, cpu
->arch
.idt
[num
].a
,
401 cpu
->arch
.idt
[num
].b
);
406 * Here's the hard part: returning to the Host every time a trap happens
407 * and then calling deliver_trap() and re-entering the Guest is slow.
408 * Particularly because Guest userspace system calls are traps (usually trap
411 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
412 * into the Guest. This is possible, but the complexities cause the size of
413 * this file to double! However, 150 lines of code is worth writing for taking
414 * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
415 * the other hypervisors would beat it up at lunchtime.
417 * This routine indicates if a particular trap number could be delivered
420 * Unfortunately, Linux 4.6 started using an interrupt gate instead of a
421 * trap gate for syscalls, so this trick is ineffective. See Mastery for
422 * how we could do this anyway...
424 static bool direct_trap(unsigned int num
)
427 * Hardware interrupts don't go to the Guest at all (except system
430 if (num
>= FIRST_EXTERNAL_VECTOR
&& !could_be_syscall(num
))
434 * The Host needs to see page faults (for shadow paging and to save the
435 * fault address), general protection faults (in/out emulation) and
436 * device not available (TS handling) and of course, the hypercall trap.
438 return num
!= 14 && num
!= 13 && num
!= 7 && num
!= LGUEST_TRAP_ENTRY
;
443 * The Guest has the ability to turn its interrupt gates into trap gates,
444 * if it is careful. The Host will let trap gates can go directly to the
445 * Guest, but the Guest needs the interrupts atomically disabled for an
446 * interrupt gate. The Host could provide a mechanism to register more
447 * "no-interrupt" regions, and the Guest could point the trap gate at
448 * instructions within that region, where it can safely disable interrupts.
452 * The Guests do not use the sysenter (fast system call) instruction,
453 * because it's hardcoded to enter privilege level 0 and so can't go direct.
454 * It's about twice as fast as the older "int 0x80" system call, so it might
455 * still be worthwhile to handle it in the Switcher and lcall down to the
456 * Guest. The sysenter semantics are hairy tho: search for that keyword in
461 * When we make traps go directly into the Guest, we need to make sure
462 * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the
463 * CPU trying to deliver the trap will fault while trying to push the interrupt
464 * words on the stack: this is called a double fault, and it forces us to kill
467 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault.
469 void pin_stack_pages(struct lg_cpu
*cpu
)
474 * Depending on the CONFIG_4KSTACKS option, the Guest can have one or
475 * two pages of stack space.
477 for (i
= 0; i
< cpu
->lg
->stack_pages
; i
++)
479 * The stack grows *upwards*, so the address we're given is the
480 * start of the page after the kernel stack. Subtract one to
481 * get back onto the first stack page, and keep subtracting to
482 * get to the rest of the stack pages.
484 pin_page(cpu
, cpu
->esp1
- 1 - i
* PAGE_SIZE
);
488 * Direct traps also mean that we need to know whenever the Guest wants to use
489 * a different kernel stack, so we can change the guest TSS to use that
490 * stack. The TSS entries expect a virtual address, so unlike most addresses
491 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
494 * In Linux each process has its own kernel stack, so this happens a lot: we
495 * change stacks on each context switch.
497 void guest_set_stack(struct lg_cpu
*cpu
, u32 seg
, u32 esp
, unsigned int pages
)
500 * You're not allowed a stack segment with privilege level 0: bad Guest!
502 if ((seg
& 0x3) != GUEST_PL
)
503 kill_guest(cpu
, "bad stack segment %i", seg
);
504 /* We only expect one or two stack pages. */
506 kill_guest(cpu
, "bad stack pages %u", pages
);
507 /* Save where the stack is, and how many pages */
510 cpu
->lg
->stack_pages
= pages
;
511 /* Make sure the new stack pages are mapped */
512 pin_stack_pages(cpu
);
516 * All this reference to mapping stacks leads us neatly into the other complex
517 * part of the Host: page table handling.
521 * This is the routine which actually checks the Guest's IDT entry and
522 * transfers it into the entry in "struct lguest":
524 static void set_trap(struct lg_cpu
*cpu
, struct desc_struct
*trap
,
525 unsigned int num
, u32 lo
, u32 hi
)
527 u8 type
= idt_type(lo
, hi
);
529 /* We zero-out a not-present entry */
530 if (!idt_present(lo
, hi
)) {
531 trap
->a
= trap
->b
= 0;
535 /* We only support interrupt and trap gates. */
536 if (type
!= 0xE && type
!= 0xF)
537 kill_guest(cpu
, "bad IDT type %i", type
);
540 * We only copy the handler address, present bit, privilege level and
541 * type. The privilege level controls where the trap can be triggered
542 * manually with an "int" instruction. This is usually GUEST_PL,
543 * except for system calls which userspace can use.
545 trap
->a
= ((__KERNEL_CS
|GUEST_PL
)<<16) | (lo
&0x0000FFFF);
546 trap
->b
= (hi
&0xFFFFEF00);
550 * While we're here, dealing with delivering traps and interrupts to the
551 * Guest, we might as well complete the picture: how the Guest tells us where
552 * it wants them to go. This would be simple, except making traps fast
553 * requires some tricks.
555 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
556 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here.
558 void load_guest_idt_entry(struct lg_cpu
*cpu
, unsigned int num
, u32 lo
, u32 hi
)
561 * Guest never handles: NMI, doublefault, spurious interrupt or
562 * hypercall. We ignore when it tries to set them.
564 if (num
== 2 || num
== 8 || num
== 15 || num
== LGUEST_TRAP_ENTRY
)
568 * Mark the IDT as changed: next time the Guest runs we'll know we have
569 * to copy this again.
571 cpu
->changed
|= CHANGED_IDT
;
573 /* Check that the Guest doesn't try to step outside the bounds. */
574 if (num
>= ARRAY_SIZE(cpu
->arch
.idt
))
575 kill_guest(cpu
, "Setting idt entry %u", num
);
577 set_trap(cpu
, &cpu
->arch
.idt
[num
], num
, lo
, hi
);
581 * The default entry for each interrupt points into the Switcher routines which
582 * simply return to the Host. The run_guest() loop will then call
583 * deliver_trap() to bounce it back into the Guest.
585 static void default_idt_entry(struct desc_struct
*idt
,
587 const unsigned long handler
,
588 const struct desc_struct
*base
)
590 /* A present interrupt gate. */
594 * Set the privilege level on the entry for the hypercall: this allows
595 * the Guest to use the "int" instruction to trigger it.
597 if (trap
== LGUEST_TRAP_ENTRY
)
598 flags
|= (GUEST_PL
<< 13);
601 * Copy privilege level from what Guest asked for. This allows
602 * debug (int 3) traps from Guest userspace, for example.
604 flags
|= (base
->b
& 0x6000);
606 /* Now pack it into the IDT entry in its weird format. */
607 idt
->a
= (LGUEST_CS
<<16) | (handler
&0x0000FFFF);
608 idt
->b
= (handler
&0xFFFF0000) | flags
;
611 /* When the Guest first starts, we put default entries into the IDT. */
612 void setup_default_idt_entries(struct lguest_ro_state
*state
,
613 const unsigned long *def
)
617 for (i
= 0; i
< ARRAY_SIZE(state
->guest_idt
); i
++)
618 default_idt_entry(&state
->guest_idt
[i
], i
, def
[i
], NULL
);
622 * We don't use the IDT entries in the "struct lguest" directly, instead
623 * we copy them into the IDT which we've set up for Guests on this CPU, just
624 * before we run the Guest. This routine does that copy.
626 void copy_traps(const struct lg_cpu
*cpu
, struct desc_struct
*idt
,
627 const unsigned long *def
)
632 * We can simply copy the direct traps, otherwise we use the default
633 * ones in the Switcher: they will return to the Host.
635 for (i
= 0; i
< ARRAY_SIZE(cpu
->arch
.idt
); i
++) {
636 const struct desc_struct
*gidt
= &cpu
->arch
.idt
[i
];
638 /* If no Guest can ever override this trap, leave it alone. */
643 * Only trap gates (type 15) can go direct to the Guest.
644 * Interrupt gates (type 14) disable interrupts as they are
645 * entered, which we never let the Guest do. Not present
646 * entries (type 0x0) also can't go direct, of course.
648 * If it can't go direct, we still need to copy the priv. level:
649 * they might want to give userspace access to a software
652 if (idt_type(gidt
->a
, gidt
->b
) == 0xF)
655 default_idt_entry(&idt
[i
], i
, def
[i
], gidt
);
662 * There are two sources of virtual interrupts. We saw one in lguest_user.c:
663 * the Launcher sending interrupts for virtual devices. The other is the Guest
666 * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
667 * the next timer interrupt (in nanoseconds). We use the high-resolution timer
668 * infrastructure to set a callback at that time.
670 * 0 means "turn off the clock".
672 void guest_set_clockevent(struct lg_cpu
*cpu
, unsigned long delta
)
676 if (unlikely(delta
== 0)) {
677 /* Clock event device is shutting down. */
678 hrtimer_cancel(&cpu
->hrt
);
683 * We use wallclock time here, so the Guest might not be running for
684 * all the time between now and the timer interrupt it asked for. This
685 * is almost always the right thing to do.
687 expires
= ktime_add_ns(ktime_get_real(), delta
);
688 hrtimer_start(&cpu
->hrt
, expires
, HRTIMER_MODE_ABS
);
691 /* This is the function called when the Guest's timer expires. */
692 static enum hrtimer_restart
clockdev_fn(struct hrtimer
*timer
)
694 struct lg_cpu
*cpu
= container_of(timer
, struct lg_cpu
, hrt
);
696 /* Remember the first interrupt is the timer interrupt. */
697 set_interrupt(cpu
, 0);
698 return HRTIMER_NORESTART
;
701 /* This sets up the timer for this Guest. */
702 void init_clockdev(struct lg_cpu
*cpu
)
704 hrtimer_init(&cpu
->hrt
, CLOCK_REALTIME
, HRTIMER_MODE_ABS
);
705 cpu
->hrt
.function
= clockdev_fn
;