Loading...
1/*P:800
2 * Interrupts (traps) are complicated enough to earn their own file.
3 * There are three classes of interrupts:
4 *
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.
8 *
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.
15:*/
16#include <linux/uaccess.h>
17#include <linux/interrupt.h>
18#include <linux/module.h>
19#include <linux/sched.h>
20#include "lg.h"
21
22/* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */
23static unsigned int syscall_vector = SYSCALL_VECTOR;
24module_param(syscall_vector, uint, 0444);
25
26/* The address of the interrupt handler is split into two bits: */
27static unsigned long idt_address(u32 lo, u32 hi)
28{
29 return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
30}
31
32/*
33 * The "type" of the interrupt handler is a 4 bit field: we only support a
34 * couple of types.
35 */
36static int idt_type(u32 lo, u32 hi)
37{
38 return (hi >> 8) & 0xF;
39}
40
41/* An IDT entry can't be used unless the "present" bit is set. */
42static bool idt_present(u32 lo, u32 hi)
43{
44 return (hi & 0x8000);
45}
46
47/*
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.
50 */
51static void push_guest_stack(struct lg_cpu *cpu, unsigned long *gstack, u32 val)
52{
53 /* Stack grows upwards: move stack then write value. */
54 *gstack -= 4;
55 lgwrite(cpu, *gstack, u32, val);
56}
57
58/*H:210
59 * The set_guest_interrupt() routine actually delivers the interrupt or
60 * trap. The mechanics of delivering traps and interrupts to the Guest are the
61 * same, except some traps have an "error code" which gets pushed onto the
62 * stack as well: the caller tells us if this is one.
63 *
64 * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
65 * interrupt or trap. It's split into two parts for traditional reasons: gcc
66 * on i386 used to be frightened by 64 bit numbers.
67 *
68 * We set up the stack just like the CPU does for a real interrupt, so it's
69 * identical for the Guest (and the standard "iret" instruction will undo
70 * it).
71 */
72static void set_guest_interrupt(struct lg_cpu *cpu, u32 lo, u32 hi,
73 bool has_err)
74{
75 unsigned long gstack, origstack;
76 u32 eflags, ss, irq_enable;
77 unsigned long virtstack;
78
79 /*
80 * There are two cases for interrupts: one where the Guest is already
81 * in the kernel, and a more complex one where the Guest is in
82 * userspace. We check the privilege level to find out.
83 */
84 if ((cpu->regs->ss&0x3) != GUEST_PL) {
85 /*
86 * The Guest told us their kernel stack with the SET_STACK
87 * hypercall: both the virtual address and the segment.
88 */
89 virtstack = cpu->esp1;
90 ss = cpu->ss1;
91
92 origstack = gstack = guest_pa(cpu, virtstack);
93 /*
94 * We push the old stack segment and pointer onto the new
95 * stack: when the Guest does an "iret" back from the interrupt
96 * handler the CPU will notice they're dropping privilege
97 * levels and expect these here.
98 */
99 push_guest_stack(cpu, &gstack, cpu->regs->ss);
100 push_guest_stack(cpu, &gstack, cpu->regs->esp);
101 } else {
102 /* We're staying on the same Guest (kernel) stack. */
103 virtstack = cpu->regs->esp;
104 ss = cpu->regs->ss;
105
106 origstack = gstack = guest_pa(cpu, virtstack);
107 }
108
109 /*
110 * Remember that we never let the Guest actually disable interrupts, so
111 * the "Interrupt Flag" bit is always set. We copy that bit from the
112 * Guest's "irq_enabled" field into the eflags word: we saw the Guest
113 * copy it back in "lguest_iret".
114 */
115 eflags = cpu->regs->eflags;
116 if (get_user(irq_enable, &cpu->lg->lguest_data->irq_enabled) == 0
117 && !(irq_enable & X86_EFLAGS_IF))
118 eflags &= ~X86_EFLAGS_IF;
119
120 /*
121 * An interrupt is expected to push three things on the stack: the old
122 * "eflags" word, the old code segment, and the old instruction
123 * pointer.
124 */
125 push_guest_stack(cpu, &gstack, eflags);
126 push_guest_stack(cpu, &gstack, cpu->regs->cs);
127 push_guest_stack(cpu, &gstack, cpu->regs->eip);
128
129 /* For the six traps which supply an error code, we push that, too. */
130 if (has_err)
131 push_guest_stack(cpu, &gstack, cpu->regs->errcode);
132
133 /*
134 * Now we've pushed all the old state, we change the stack, the code
135 * segment and the address to execute.
136 */
137 cpu->regs->ss = ss;
138 cpu->regs->esp = virtstack + (gstack - origstack);
139 cpu->regs->cs = (__KERNEL_CS|GUEST_PL);
140 cpu->regs->eip = idt_address(lo, hi);
141
142 /*
143 * There are two kinds of interrupt handlers: 0xE is an "interrupt
144 * gate" which expects interrupts to be disabled on entry.
145 */
146 if (idt_type(lo, hi) == 0xE)
147 if (put_user(0, &cpu->lg->lguest_data->irq_enabled))
148 kill_guest(cpu, "Disabling interrupts");
149}
150
151/*H:205
152 * Virtual Interrupts.
153 *
154 * interrupt_pending() returns the first pending interrupt which isn't blocked
155 * by the Guest. It is called before every entry to the Guest, and just before
156 * we go to sleep when the Guest has halted itself.
157 */
158unsigned int interrupt_pending(struct lg_cpu *cpu, bool *more)
159{
160 unsigned int irq;
161 DECLARE_BITMAP(blk, LGUEST_IRQS);
162
163 /* If the Guest hasn't even initialized yet, we can do nothing. */
164 if (!cpu->lg->lguest_data)
165 return LGUEST_IRQS;
166
167 /*
168 * Take our "irqs_pending" array and remove any interrupts the Guest
169 * wants blocked: the result ends up in "blk".
170 */
171 if (copy_from_user(&blk, cpu->lg->lguest_data->blocked_interrupts,
172 sizeof(blk)))
173 return LGUEST_IRQS;
174 bitmap_andnot(blk, cpu->irqs_pending, blk, LGUEST_IRQS);
175
176 /* Find the first interrupt. */
177 irq = find_first_bit(blk, LGUEST_IRQS);
178 *more = find_next_bit(blk, LGUEST_IRQS, irq+1);
179
180 return irq;
181}
182
183/*
184 * This actually diverts the Guest to running an interrupt handler, once an
185 * interrupt has been identified by interrupt_pending().
186 */
187void try_deliver_interrupt(struct lg_cpu *cpu, unsigned int irq, bool more)
188{
189 struct desc_struct *idt;
190
191 BUG_ON(irq >= LGUEST_IRQS);
192
193 /*
194 * They may be in the middle of an iret, where they asked us never to
195 * deliver interrupts.
196 */
197 if (cpu->regs->eip >= cpu->lg->noirq_start &&
198 (cpu->regs->eip < cpu->lg->noirq_end))
199 return;
200
201 /* If they're halted, interrupts restart them. */
202 if (cpu->halted) {
203 /* Re-enable interrupts. */
204 if (put_user(X86_EFLAGS_IF, &cpu->lg->lguest_data->irq_enabled))
205 kill_guest(cpu, "Re-enabling interrupts");
206 cpu->halted = 0;
207 } else {
208 /* Otherwise we check if they have interrupts disabled. */
209 u32 irq_enabled;
210 if (get_user(irq_enabled, &cpu->lg->lguest_data->irq_enabled))
211 irq_enabled = 0;
212 if (!irq_enabled) {
213 /* Make sure they know an IRQ is pending. */
214 put_user(X86_EFLAGS_IF,
215 &cpu->lg->lguest_data->irq_pending);
216 return;
217 }
218 }
219
220 /*
221 * Look at the IDT entry the Guest gave us for this interrupt. The
222 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
223 * over them.
224 */
225 idt = &cpu->arch.idt[FIRST_EXTERNAL_VECTOR+irq];
226 /* If they don't have a handler (yet?), we just ignore it */
227 if (idt_present(idt->a, idt->b)) {
228 /* OK, mark it no longer pending and deliver it. */
229 clear_bit(irq, cpu->irqs_pending);
230 /*
231 * set_guest_interrupt() takes the interrupt descriptor and a
232 * flag to say whether this interrupt pushes an error code onto
233 * the stack as well: virtual interrupts never do.
234 */
235 set_guest_interrupt(cpu, idt->a, idt->b, false);
236 }
237
238 /*
239 * Every time we deliver an interrupt, we update the timestamp in the
240 * Guest's lguest_data struct. It would be better for the Guest if we
241 * did this more often, but it can actually be quite slow: doing it
242 * here is a compromise which means at least it gets updated every
243 * timer interrupt.
244 */
245 write_timestamp(cpu);
246
247 /*
248 * If there are no other interrupts we want to deliver, clear
249 * the pending flag.
250 */
251 if (!more)
252 put_user(0, &cpu->lg->lguest_data->irq_pending);
253}
254
255/* And this is the routine when we want to set an interrupt for the Guest. */
256void set_interrupt(struct lg_cpu *cpu, unsigned int irq)
257{
258 /*
259 * Next time the Guest runs, the core code will see if it can deliver
260 * this interrupt.
261 */
262 set_bit(irq, cpu->irqs_pending);
263
264 /*
265 * Make sure it sees it; it might be asleep (eg. halted), or running
266 * the Guest right now, in which case kick_process() will knock it out.
267 */
268 if (!wake_up_process(cpu->tsk))
269 kick_process(cpu->tsk);
270}
271/*:*/
272
273/*
274 * Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent
275 * me a patch, so we support that too. It'd be a big step for lguest if half
276 * the Plan 9 user base were to start using it.
277 *
278 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
279 * userbase. Oh well.
280 */
281static bool could_be_syscall(unsigned int num)
282{
283 /* Normal Linux SYSCALL_VECTOR or reserved vector? */
284 return num == SYSCALL_VECTOR || num == syscall_vector;
285}
286
287/* The syscall vector it wants must be unused by Host. */
288bool check_syscall_vector(struct lguest *lg)
289{
290 u32 vector;
291
292 if (get_user(vector, &lg->lguest_data->syscall_vec))
293 return false;
294
295 return could_be_syscall(vector);
296}
297
298int init_interrupts(void)
299{
300 /* If they want some strange system call vector, reserve it now */
301 if (syscall_vector != SYSCALL_VECTOR) {
302 if (test_bit(syscall_vector, used_vectors) ||
303 vector_used_by_percpu_irq(syscall_vector)) {
304 printk(KERN_ERR "lg: couldn't reserve syscall %u\n",
305 syscall_vector);
306 return -EBUSY;
307 }
308 set_bit(syscall_vector, used_vectors);
309 }
310
311 return 0;
312}
313
314void free_interrupts(void)
315{
316 if (syscall_vector != SYSCALL_VECTOR)
317 clear_bit(syscall_vector, used_vectors);
318}
319
320/*H:220
321 * Now we've got the routines to deliver interrupts, delivering traps like
322 * page fault is easy. The only trick is that Intel decided that some traps
323 * should have error codes:
324 */
325static bool has_err(unsigned int trap)
326{
327 return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
328}
329
330/* deliver_trap() returns true if it could deliver the trap. */
331bool deliver_trap(struct lg_cpu *cpu, unsigned int num)
332{
333 /*
334 * Trap numbers are always 8 bit, but we set an impossible trap number
335 * for traps inside the Switcher, so check that here.
336 */
337 if (num >= ARRAY_SIZE(cpu->arch.idt))
338 return false;
339
340 /*
341 * Early on the Guest hasn't set the IDT entries (or maybe it put a
342 * bogus one in): if we fail here, the Guest will be killed.
343 */
344 if (!idt_present(cpu->arch.idt[num].a, cpu->arch.idt[num].b))
345 return false;
346 set_guest_interrupt(cpu, cpu->arch.idt[num].a,
347 cpu->arch.idt[num].b, has_err(num));
348 return true;
349}
350
351/*H:250
352 * Here's the hard part: returning to the Host every time a trap happens
353 * and then calling deliver_trap() and re-entering the Guest is slow.
354 * Particularly because Guest userspace system calls are traps (usually trap
355 * 128).
356 *
357 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
358 * into the Guest. This is possible, but the complexities cause the size of
359 * this file to double! However, 150 lines of code is worth writing for taking
360 * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
361 * the other hypervisors would beat it up at lunchtime.
362 *
363 * This routine indicates if a particular trap number could be delivered
364 * directly.
365 */
366static bool direct_trap(unsigned int num)
367{
368 /*
369 * Hardware interrupts don't go to the Guest at all (except system
370 * call).
371 */
372 if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
373 return false;
374
375 /*
376 * The Host needs to see page faults (for shadow paging and to save the
377 * fault address), general protection faults (in/out emulation) and
378 * device not available (TS handling) and of course, the hypercall trap.
379 */
380 return num != 14 && num != 13 && num != 7 && num != LGUEST_TRAP_ENTRY;
381}
382/*:*/
383
384/*M:005
385 * The Guest has the ability to turn its interrupt gates into trap gates,
386 * if it is careful. The Host will let trap gates can go directly to the
387 * Guest, but the Guest needs the interrupts atomically disabled for an
388 * interrupt gate. It can do this by pointing the trap gate at instructions
389 * within noirq_start and noirq_end, where it can safely disable interrupts.
390 */
391
392/*M:006
393 * The Guests do not use the sysenter (fast system call) instruction,
394 * because it's hardcoded to enter privilege level 0 and so can't go direct.
395 * It's about twice as fast as the older "int 0x80" system call, so it might
396 * still be worthwhile to handle it in the Switcher and lcall down to the
397 * Guest. The sysenter semantics are hairy tho: search for that keyword in
398 * entry.S
399:*/
400
401/*H:260
402 * When we make traps go directly into the Guest, we need to make sure
403 * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the
404 * CPU trying to deliver the trap will fault while trying to push the interrupt
405 * words on the stack: this is called a double fault, and it forces us to kill
406 * the Guest.
407 *
408 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault.
409 */
410void pin_stack_pages(struct lg_cpu *cpu)
411{
412 unsigned int i;
413
414 /*
415 * Depending on the CONFIG_4KSTACKS option, the Guest can have one or
416 * two pages of stack space.
417 */
418 for (i = 0; i < cpu->lg->stack_pages; i++)
419 /*
420 * The stack grows *upwards*, so the address we're given is the
421 * start of the page after the kernel stack. Subtract one to
422 * get back onto the first stack page, and keep subtracting to
423 * get to the rest of the stack pages.
424 */
425 pin_page(cpu, cpu->esp1 - 1 - i * PAGE_SIZE);
426}
427
428/*
429 * Direct traps also mean that we need to know whenever the Guest wants to use
430 * a different kernel stack, so we can change the guest TSS to use that
431 * stack. The TSS entries expect a virtual address, so unlike most addresses
432 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
433 * physical.
434 *
435 * In Linux each process has its own kernel stack, so this happens a lot: we
436 * change stacks on each context switch.
437 */
438void guest_set_stack(struct lg_cpu *cpu, u32 seg, u32 esp, unsigned int pages)
439{
440 /*
441 * You're not allowed a stack segment with privilege level 0: bad Guest!
442 */
443 if ((seg & 0x3) != GUEST_PL)
444 kill_guest(cpu, "bad stack segment %i", seg);
445 /* We only expect one or two stack pages. */
446 if (pages > 2)
447 kill_guest(cpu, "bad stack pages %u", pages);
448 /* Save where the stack is, and how many pages */
449 cpu->ss1 = seg;
450 cpu->esp1 = esp;
451 cpu->lg->stack_pages = pages;
452 /* Make sure the new stack pages are mapped */
453 pin_stack_pages(cpu);
454}
455
456/*
457 * All this reference to mapping stacks leads us neatly into the other complex
458 * part of the Host: page table handling.
459 */
460
461/*H:235
462 * This is the routine which actually checks the Guest's IDT entry and
463 * transfers it into the entry in "struct lguest":
464 */
465static void set_trap(struct lg_cpu *cpu, struct desc_struct *trap,
466 unsigned int num, u32 lo, u32 hi)
467{
468 u8 type = idt_type(lo, hi);
469
470 /* We zero-out a not-present entry */
471 if (!idt_present(lo, hi)) {
472 trap->a = trap->b = 0;
473 return;
474 }
475
476 /* We only support interrupt and trap gates. */
477 if (type != 0xE && type != 0xF)
478 kill_guest(cpu, "bad IDT type %i", type);
479
480 /*
481 * We only copy the handler address, present bit, privilege level and
482 * type. The privilege level controls where the trap can be triggered
483 * manually with an "int" instruction. This is usually GUEST_PL,
484 * except for system calls which userspace can use.
485 */
486 trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
487 trap->b = (hi&0xFFFFEF00);
488}
489
490/*H:230
491 * While we're here, dealing with delivering traps and interrupts to the
492 * Guest, we might as well complete the picture: how the Guest tells us where
493 * it wants them to go. This would be simple, except making traps fast
494 * requires some tricks.
495 *
496 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
497 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here.
498 */
499void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int num, u32 lo, u32 hi)
500{
501 /*
502 * Guest never handles: NMI, doublefault, spurious interrupt or
503 * hypercall. We ignore when it tries to set them.
504 */
505 if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
506 return;
507
508 /*
509 * Mark the IDT as changed: next time the Guest runs we'll know we have
510 * to copy this again.
511 */
512 cpu->changed |= CHANGED_IDT;
513
514 /* Check that the Guest doesn't try to step outside the bounds. */
515 if (num >= ARRAY_SIZE(cpu->arch.idt))
516 kill_guest(cpu, "Setting idt entry %u", num);
517 else
518 set_trap(cpu, &cpu->arch.idt[num], num, lo, hi);
519}
520
521/*
522 * The default entry for each interrupt points into the Switcher routines which
523 * simply return to the Host. The run_guest() loop will then call
524 * deliver_trap() to bounce it back into the Guest.
525 */
526static void default_idt_entry(struct desc_struct *idt,
527 int trap,
528 const unsigned long handler,
529 const struct desc_struct *base)
530{
531 /* A present interrupt gate. */
532 u32 flags = 0x8e00;
533
534 /*
535 * Set the privilege level on the entry for the hypercall: this allows
536 * the Guest to use the "int" instruction to trigger it.
537 */
538 if (trap == LGUEST_TRAP_ENTRY)
539 flags |= (GUEST_PL << 13);
540 else if (base)
541 /*
542 * Copy privilege level from what Guest asked for. This allows
543 * debug (int 3) traps from Guest userspace, for example.
544 */
545 flags |= (base->b & 0x6000);
546
547 /* Now pack it into the IDT entry in its weird format. */
548 idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
549 idt->b = (handler&0xFFFF0000) | flags;
550}
551
552/* When the Guest first starts, we put default entries into the IDT. */
553void setup_default_idt_entries(struct lguest_ro_state *state,
554 const unsigned long *def)
555{
556 unsigned int i;
557
558 for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
559 default_idt_entry(&state->guest_idt[i], i, def[i], NULL);
560}
561
562/*H:240
563 * We don't use the IDT entries in the "struct lguest" directly, instead
564 * we copy them into the IDT which we've set up for Guests on this CPU, just
565 * before we run the Guest. This routine does that copy.
566 */
567void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt,
568 const unsigned long *def)
569{
570 unsigned int i;
571
572 /*
573 * We can simply copy the direct traps, otherwise we use the default
574 * ones in the Switcher: they will return to the Host.
575 */
576 for (i = 0; i < ARRAY_SIZE(cpu->arch.idt); i++) {
577 const struct desc_struct *gidt = &cpu->arch.idt[i];
578
579 /* If no Guest can ever override this trap, leave it alone. */
580 if (!direct_trap(i))
581 continue;
582
583 /*
584 * Only trap gates (type 15) can go direct to the Guest.
585 * Interrupt gates (type 14) disable interrupts as they are
586 * entered, which we never let the Guest do. Not present
587 * entries (type 0x0) also can't go direct, of course.
588 *
589 * If it can't go direct, we still need to copy the priv. level:
590 * they might want to give userspace access to a software
591 * interrupt.
592 */
593 if (idt_type(gidt->a, gidt->b) == 0xF)
594 idt[i] = *gidt;
595 else
596 default_idt_entry(&idt[i], i, def[i], gidt);
597 }
598}
599
600/*H:200
601 * The Guest Clock.
602 *
603 * There are two sources of virtual interrupts. We saw one in lguest_user.c:
604 * the Launcher sending interrupts for virtual devices. The other is the Guest
605 * timer interrupt.
606 *
607 * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
608 * the next timer interrupt (in nanoseconds). We use the high-resolution timer
609 * infrastructure to set a callback at that time.
610 *
611 * 0 means "turn off the clock".
612 */
613void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta)
614{
615 ktime_t expires;
616
617 if (unlikely(delta == 0)) {
618 /* Clock event device is shutting down. */
619 hrtimer_cancel(&cpu->hrt);
620 return;
621 }
622
623 /*
624 * We use wallclock time here, so the Guest might not be running for
625 * all the time between now and the timer interrupt it asked for. This
626 * is almost always the right thing to do.
627 */
628 expires = ktime_add_ns(ktime_get_real(), delta);
629 hrtimer_start(&cpu->hrt, expires, HRTIMER_MODE_ABS);
630}
631
632/* This is the function called when the Guest's timer expires. */
633static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
634{
635 struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt);
636
637 /* Remember the first interrupt is the timer interrupt. */
638 set_interrupt(cpu, 0);
639 return HRTIMER_NORESTART;
640}
641
642/* This sets up the timer for this Guest. */
643void init_clockdev(struct lg_cpu *cpu)
644{
645 hrtimer_init(&cpu->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
646 cpu->hrt.function = clockdev_fn;
647}
1/*P:800
2 * Interrupts (traps) are complicated enough to earn their own file.
3 * There are three classes of interrupts:
4 *
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.
8 *
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.
15:*/
16#include <linux/uaccess.h>
17#include <linux/interrupt.h>
18#include <linux/module.h>
19#include <linux/sched.h>
20#include "lg.h"
21
22/* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */
23static unsigned int syscall_vector = SYSCALL_VECTOR;
24module_param(syscall_vector, uint, 0444);
25
26/* The address of the interrupt handler is split into two bits: */
27static unsigned long idt_address(u32 lo, u32 hi)
28{
29 return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
30}
31
32/*
33 * The "type" of the interrupt handler is a 4 bit field: we only support a
34 * couple of types.
35 */
36static int idt_type(u32 lo, u32 hi)
37{
38 return (hi >> 8) & 0xF;
39}
40
41/* An IDT entry can't be used unless the "present" bit is set. */
42static bool idt_present(u32 lo, u32 hi)
43{
44 return (hi & 0x8000);
45}
46
47/*
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.
50 */
51static void push_guest_stack(struct lg_cpu *cpu, unsigned long *gstack, u32 val)
52{
53 /* Stack grows upwards: move stack then write value. */
54 *gstack -= 4;
55 lgwrite(cpu, *gstack, u32, val);
56}
57
58/*H:210
59 * The set_guest_interrupt() routine actually delivers the interrupt or
60 * trap. The mechanics of delivering traps and interrupts to the Guest are the
61 * same, except some traps have an "error code" which gets pushed onto the
62 * stack as well: the caller tells us if this is one.
63 *
64 * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
65 * interrupt or trap. It's split into two parts for traditional reasons: gcc
66 * on i386 used to be frightened by 64 bit numbers.
67 *
68 * We set up the stack just like the CPU does for a real interrupt, so it's
69 * identical for the Guest (and the standard "iret" instruction will undo
70 * it).
71 */
72static void set_guest_interrupt(struct lg_cpu *cpu, u32 lo, u32 hi,
73 bool has_err)
74{
75 unsigned long gstack, origstack;
76 u32 eflags, ss, irq_enable;
77 unsigned long virtstack;
78
79 /*
80 * There are two cases for interrupts: one where the Guest is already
81 * in the kernel, and a more complex one where the Guest is in
82 * userspace. We check the privilege level to find out.
83 */
84 if ((cpu->regs->ss&0x3) != GUEST_PL) {
85 /*
86 * The Guest told us their kernel stack with the SET_STACK
87 * hypercall: both the virtual address and the segment.
88 */
89 virtstack = cpu->esp1;
90 ss = cpu->ss1;
91
92 origstack = gstack = guest_pa(cpu, virtstack);
93 /*
94 * We push the old stack segment and pointer onto the new
95 * stack: when the Guest does an "iret" back from the interrupt
96 * handler the CPU will notice they're dropping privilege
97 * levels and expect these here.
98 */
99 push_guest_stack(cpu, &gstack, cpu->regs->ss);
100 push_guest_stack(cpu, &gstack, cpu->regs->esp);
101 } else {
102 /* We're staying on the same Guest (kernel) stack. */
103 virtstack = cpu->regs->esp;
104 ss = cpu->regs->ss;
105
106 origstack = gstack = guest_pa(cpu, virtstack);
107 }
108
109 /*
110 * Remember that we never let the Guest actually disable interrupts, so
111 * the "Interrupt Flag" bit is always set. We copy that bit from the
112 * Guest's "irq_enabled" field into the eflags word: we saw the Guest
113 * copy it back in "lguest_iret".
114 */
115 eflags = cpu->regs->eflags;
116 if (get_user(irq_enable, &cpu->lg->lguest_data->irq_enabled) == 0
117 && !(irq_enable & X86_EFLAGS_IF))
118 eflags &= ~X86_EFLAGS_IF;
119
120 /*
121 * An interrupt is expected to push three things on the stack: the old
122 * "eflags" word, the old code segment, and the old instruction
123 * pointer.
124 */
125 push_guest_stack(cpu, &gstack, eflags);
126 push_guest_stack(cpu, &gstack, cpu->regs->cs);
127 push_guest_stack(cpu, &gstack, cpu->regs->eip);
128
129 /* For the six traps which supply an error code, we push that, too. */
130 if (has_err)
131 push_guest_stack(cpu, &gstack, cpu->regs->errcode);
132
133 /*
134 * Now we've pushed all the old state, we change the stack, the code
135 * segment and the address to execute.
136 */
137 cpu->regs->ss = ss;
138 cpu->regs->esp = virtstack + (gstack - origstack);
139 cpu->regs->cs = (__KERNEL_CS|GUEST_PL);
140 cpu->regs->eip = idt_address(lo, hi);
141
142 /*
143 * Trapping always clears these flags:
144 * TF: Trap flag
145 * VM: Virtual 8086 mode
146 * RF: Resume
147 * NT: Nested task.
148 */
149 cpu->regs->eflags &=
150 ~(X86_EFLAGS_TF|X86_EFLAGS_VM|X86_EFLAGS_RF|X86_EFLAGS_NT);
151
152 /*
153 * There are two kinds of interrupt handlers: 0xE is an "interrupt
154 * gate" which expects interrupts to be disabled on entry.
155 */
156 if (idt_type(lo, hi) == 0xE)
157 if (put_user(0, &cpu->lg->lguest_data->irq_enabled))
158 kill_guest(cpu, "Disabling interrupts");
159}
160
161/*H:205
162 * Virtual Interrupts.
163 *
164 * interrupt_pending() returns the first pending interrupt which isn't blocked
165 * by the Guest. It is called before every entry to the Guest, and just before
166 * we go to sleep when the Guest has halted itself.
167 */
168unsigned int interrupt_pending(struct lg_cpu *cpu, bool *more)
169{
170 unsigned int irq;
171 DECLARE_BITMAP(blk, LGUEST_IRQS);
172
173 /* If the Guest hasn't even initialized yet, we can do nothing. */
174 if (!cpu->lg->lguest_data)
175 return LGUEST_IRQS;
176
177 /*
178 * Take our "irqs_pending" array and remove any interrupts the Guest
179 * wants blocked: the result ends up in "blk".
180 */
181 if (copy_from_user(&blk, cpu->lg->lguest_data->blocked_interrupts,
182 sizeof(blk)))
183 return LGUEST_IRQS;
184 bitmap_andnot(blk, cpu->irqs_pending, blk, LGUEST_IRQS);
185
186 /* Find the first interrupt. */
187 irq = find_first_bit(blk, LGUEST_IRQS);
188 *more = find_next_bit(blk, LGUEST_IRQS, irq+1);
189
190 return irq;
191}
192
193/*
194 * This actually diverts the Guest to running an interrupt handler, once an
195 * interrupt has been identified by interrupt_pending().
196 */
197void try_deliver_interrupt(struct lg_cpu *cpu, unsigned int irq, bool more)
198{
199 struct desc_struct *idt;
200
201 BUG_ON(irq >= LGUEST_IRQS);
202
203 /*
204 * They may be in the middle of an iret, where they asked us never to
205 * deliver interrupts.
206 */
207 if (cpu->regs->eip >= cpu->lg->noirq_start &&
208 (cpu->regs->eip < cpu->lg->noirq_end))
209 return;
210
211 /* If they're halted, interrupts restart them. */
212 if (cpu->halted) {
213 /* Re-enable interrupts. */
214 if (put_user(X86_EFLAGS_IF, &cpu->lg->lguest_data->irq_enabled))
215 kill_guest(cpu, "Re-enabling interrupts");
216 cpu->halted = 0;
217 } else {
218 /* Otherwise we check if they have interrupts disabled. */
219 u32 irq_enabled;
220 if (get_user(irq_enabled, &cpu->lg->lguest_data->irq_enabled))
221 irq_enabled = 0;
222 if (!irq_enabled) {
223 /* Make sure they know an IRQ is pending. */
224 put_user(X86_EFLAGS_IF,
225 &cpu->lg->lguest_data->irq_pending);
226 return;
227 }
228 }
229
230 /*
231 * Look at the IDT entry the Guest gave us for this interrupt. The
232 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
233 * over them.
234 */
235 idt = &cpu->arch.idt[FIRST_EXTERNAL_VECTOR+irq];
236 /* If they don't have a handler (yet?), we just ignore it */
237 if (idt_present(idt->a, idt->b)) {
238 /* OK, mark it no longer pending and deliver it. */
239 clear_bit(irq, cpu->irqs_pending);
240 /*
241 * set_guest_interrupt() takes the interrupt descriptor and a
242 * flag to say whether this interrupt pushes an error code onto
243 * the stack as well: virtual interrupts never do.
244 */
245 set_guest_interrupt(cpu, idt->a, idt->b, false);
246 }
247
248 /*
249 * Every time we deliver an interrupt, we update the timestamp in the
250 * Guest's lguest_data struct. It would be better for the Guest if we
251 * did this more often, but it can actually be quite slow: doing it
252 * here is a compromise which means at least it gets updated every
253 * timer interrupt.
254 */
255 write_timestamp(cpu);
256
257 /*
258 * If there are no other interrupts we want to deliver, clear
259 * the pending flag.
260 */
261 if (!more)
262 put_user(0, &cpu->lg->lguest_data->irq_pending);
263}
264
265/* And this is the routine when we want to set an interrupt for the Guest. */
266void set_interrupt(struct lg_cpu *cpu, unsigned int irq)
267{
268 /*
269 * Next time the Guest runs, the core code will see if it can deliver
270 * this interrupt.
271 */
272 set_bit(irq, cpu->irqs_pending);
273
274 /*
275 * Make sure it sees it; it might be asleep (eg. halted), or running
276 * the Guest right now, in which case kick_process() will knock it out.
277 */
278 if (!wake_up_process(cpu->tsk))
279 kick_process(cpu->tsk);
280}
281/*:*/
282
283/*
284 * Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent
285 * me a patch, so we support that too. It'd be a big step for lguest if half
286 * the Plan 9 user base were to start using it.
287 *
288 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
289 * userbase. Oh well.
290 */
291static bool could_be_syscall(unsigned int num)
292{
293 /* Normal Linux SYSCALL_VECTOR or reserved vector? */
294 return num == SYSCALL_VECTOR || num == syscall_vector;
295}
296
297/* The syscall vector it wants must be unused by Host. */
298bool check_syscall_vector(struct lguest *lg)
299{
300 u32 vector;
301
302 if (get_user(vector, &lg->lguest_data->syscall_vec))
303 return false;
304
305 return could_be_syscall(vector);
306}
307
308int init_interrupts(void)
309{
310 /* If they want some strange system call vector, reserve it now */
311 if (syscall_vector != SYSCALL_VECTOR) {
312 if (test_bit(syscall_vector, used_vectors) ||
313 vector_used_by_percpu_irq(syscall_vector)) {
314 printk(KERN_ERR "lg: couldn't reserve syscall %u\n",
315 syscall_vector);
316 return -EBUSY;
317 }
318 set_bit(syscall_vector, used_vectors);
319 }
320
321 return 0;
322}
323
324void free_interrupts(void)
325{
326 if (syscall_vector != SYSCALL_VECTOR)
327 clear_bit(syscall_vector, used_vectors);
328}
329
330/*H:220
331 * Now we've got the routines to deliver interrupts, delivering traps like
332 * page fault is easy. The only trick is that Intel decided that some traps
333 * should have error codes:
334 */
335static bool has_err(unsigned int trap)
336{
337 return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
338}
339
340/* deliver_trap() returns true if it could deliver the trap. */
341bool deliver_trap(struct lg_cpu *cpu, unsigned int num)
342{
343 /*
344 * Trap numbers are always 8 bit, but we set an impossible trap number
345 * for traps inside the Switcher, so check that here.
346 */
347 if (num >= ARRAY_SIZE(cpu->arch.idt))
348 return false;
349
350 /*
351 * Early on the Guest hasn't set the IDT entries (or maybe it put a
352 * bogus one in): if we fail here, the Guest will be killed.
353 */
354 if (!idt_present(cpu->arch.idt[num].a, cpu->arch.idt[num].b))
355 return false;
356 set_guest_interrupt(cpu, cpu->arch.idt[num].a,
357 cpu->arch.idt[num].b, has_err(num));
358 return true;
359}
360
361/*H:250
362 * Here's the hard part: returning to the Host every time a trap happens
363 * and then calling deliver_trap() and re-entering the Guest is slow.
364 * Particularly because Guest userspace system calls are traps (usually trap
365 * 128).
366 *
367 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
368 * into the Guest. This is possible, but the complexities cause the size of
369 * this file to double! However, 150 lines of code is worth writing for taking
370 * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
371 * the other hypervisors would beat it up at lunchtime.
372 *
373 * This routine indicates if a particular trap number could be delivered
374 * directly.
375 */
376static bool direct_trap(unsigned int num)
377{
378 /*
379 * Hardware interrupts don't go to the Guest at all (except system
380 * call).
381 */
382 if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
383 return false;
384
385 /*
386 * The Host needs to see page faults (for shadow paging and to save the
387 * fault address), general protection faults (in/out emulation) and
388 * device not available (TS handling) and of course, the hypercall trap.
389 */
390 return num != 14 && num != 13 && num != 7 && num != LGUEST_TRAP_ENTRY;
391}
392/*:*/
393
394/*M:005
395 * The Guest has the ability to turn its interrupt gates into trap gates,
396 * if it is careful. The Host will let trap gates can go directly to the
397 * Guest, but the Guest needs the interrupts atomically disabled for an
398 * interrupt gate. It can do this by pointing the trap gate at instructions
399 * within noirq_start and noirq_end, where it can safely disable interrupts.
400 */
401
402/*M:006
403 * The Guests do not use the sysenter (fast system call) instruction,
404 * because it's hardcoded to enter privilege level 0 and so can't go direct.
405 * It's about twice as fast as the older "int 0x80" system call, so it might
406 * still be worthwhile to handle it in the Switcher and lcall down to the
407 * Guest. The sysenter semantics are hairy tho: search for that keyword in
408 * entry.S
409:*/
410
411/*H:260
412 * When we make traps go directly into the Guest, we need to make sure
413 * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the
414 * CPU trying to deliver the trap will fault while trying to push the interrupt
415 * words on the stack: this is called a double fault, and it forces us to kill
416 * the Guest.
417 *
418 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault.
419 */
420void pin_stack_pages(struct lg_cpu *cpu)
421{
422 unsigned int i;
423
424 /*
425 * Depending on the CONFIG_4KSTACKS option, the Guest can have one or
426 * two pages of stack space.
427 */
428 for (i = 0; i < cpu->lg->stack_pages; i++)
429 /*
430 * The stack grows *upwards*, so the address we're given is the
431 * start of the page after the kernel stack. Subtract one to
432 * get back onto the first stack page, and keep subtracting to
433 * get to the rest of the stack pages.
434 */
435 pin_page(cpu, cpu->esp1 - 1 - i * PAGE_SIZE);
436}
437
438/*
439 * Direct traps also mean that we need to know whenever the Guest wants to use
440 * a different kernel stack, so we can change the guest TSS to use that
441 * stack. The TSS entries expect a virtual address, so unlike most addresses
442 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
443 * physical.
444 *
445 * In Linux each process has its own kernel stack, so this happens a lot: we
446 * change stacks on each context switch.
447 */
448void guest_set_stack(struct lg_cpu *cpu, u32 seg, u32 esp, unsigned int pages)
449{
450 /*
451 * You're not allowed a stack segment with privilege level 0: bad Guest!
452 */
453 if ((seg & 0x3) != GUEST_PL)
454 kill_guest(cpu, "bad stack segment %i", seg);
455 /* We only expect one or two stack pages. */
456 if (pages > 2)
457 kill_guest(cpu, "bad stack pages %u", pages);
458 /* Save where the stack is, and how many pages */
459 cpu->ss1 = seg;
460 cpu->esp1 = esp;
461 cpu->lg->stack_pages = pages;
462 /* Make sure the new stack pages are mapped */
463 pin_stack_pages(cpu);
464}
465
466/*
467 * All this reference to mapping stacks leads us neatly into the other complex
468 * part of the Host: page table handling.
469 */
470
471/*H:235
472 * This is the routine which actually checks the Guest's IDT entry and
473 * transfers it into the entry in "struct lguest":
474 */
475static void set_trap(struct lg_cpu *cpu, struct desc_struct *trap,
476 unsigned int num, u32 lo, u32 hi)
477{
478 u8 type = idt_type(lo, hi);
479
480 /* We zero-out a not-present entry */
481 if (!idt_present(lo, hi)) {
482 trap->a = trap->b = 0;
483 return;
484 }
485
486 /* We only support interrupt and trap gates. */
487 if (type != 0xE && type != 0xF)
488 kill_guest(cpu, "bad IDT type %i", type);
489
490 /*
491 * We only copy the handler address, present bit, privilege level and
492 * type. The privilege level controls where the trap can be triggered
493 * manually with an "int" instruction. This is usually GUEST_PL,
494 * except for system calls which userspace can use.
495 */
496 trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
497 trap->b = (hi&0xFFFFEF00);
498}
499
500/*H:230
501 * While we're here, dealing with delivering traps and interrupts to the
502 * Guest, we might as well complete the picture: how the Guest tells us where
503 * it wants them to go. This would be simple, except making traps fast
504 * requires some tricks.
505 *
506 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
507 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here.
508 */
509void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int num, u32 lo, u32 hi)
510{
511 /*
512 * Guest never handles: NMI, doublefault, spurious interrupt or
513 * hypercall. We ignore when it tries to set them.
514 */
515 if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
516 return;
517
518 /*
519 * Mark the IDT as changed: next time the Guest runs we'll know we have
520 * to copy this again.
521 */
522 cpu->changed |= CHANGED_IDT;
523
524 /* Check that the Guest doesn't try to step outside the bounds. */
525 if (num >= ARRAY_SIZE(cpu->arch.idt))
526 kill_guest(cpu, "Setting idt entry %u", num);
527 else
528 set_trap(cpu, &cpu->arch.idt[num], num, lo, hi);
529}
530
531/*
532 * The default entry for each interrupt points into the Switcher routines which
533 * simply return to the Host. The run_guest() loop will then call
534 * deliver_trap() to bounce it back into the Guest.
535 */
536static void default_idt_entry(struct desc_struct *idt,
537 int trap,
538 const unsigned long handler,
539 const struct desc_struct *base)
540{
541 /* A present interrupt gate. */
542 u32 flags = 0x8e00;
543
544 /*
545 * Set the privilege level on the entry for the hypercall: this allows
546 * the Guest to use the "int" instruction to trigger it.
547 */
548 if (trap == LGUEST_TRAP_ENTRY)
549 flags |= (GUEST_PL << 13);
550 else if (base)
551 /*
552 * Copy privilege level from what Guest asked for. This allows
553 * debug (int 3) traps from Guest userspace, for example.
554 */
555 flags |= (base->b & 0x6000);
556
557 /* Now pack it into the IDT entry in its weird format. */
558 idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
559 idt->b = (handler&0xFFFF0000) | flags;
560}
561
562/* When the Guest first starts, we put default entries into the IDT. */
563void setup_default_idt_entries(struct lguest_ro_state *state,
564 const unsigned long *def)
565{
566 unsigned int i;
567
568 for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
569 default_idt_entry(&state->guest_idt[i], i, def[i], NULL);
570}
571
572/*H:240
573 * We don't use the IDT entries in the "struct lguest" directly, instead
574 * we copy them into the IDT which we've set up for Guests on this CPU, just
575 * before we run the Guest. This routine does that copy.
576 */
577void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt,
578 const unsigned long *def)
579{
580 unsigned int i;
581
582 /*
583 * We can simply copy the direct traps, otherwise we use the default
584 * ones in the Switcher: they will return to the Host.
585 */
586 for (i = 0; i < ARRAY_SIZE(cpu->arch.idt); i++) {
587 const struct desc_struct *gidt = &cpu->arch.idt[i];
588
589 /* If no Guest can ever override this trap, leave it alone. */
590 if (!direct_trap(i))
591 continue;
592
593 /*
594 * Only trap gates (type 15) can go direct to the Guest.
595 * Interrupt gates (type 14) disable interrupts as they are
596 * entered, which we never let the Guest do. Not present
597 * entries (type 0x0) also can't go direct, of course.
598 *
599 * If it can't go direct, we still need to copy the priv. level:
600 * they might want to give userspace access to a software
601 * interrupt.
602 */
603 if (idt_type(gidt->a, gidt->b) == 0xF)
604 idt[i] = *gidt;
605 else
606 default_idt_entry(&idt[i], i, def[i], gidt);
607 }
608}
609
610/*H:200
611 * The Guest Clock.
612 *
613 * There are two sources of virtual interrupts. We saw one in lguest_user.c:
614 * the Launcher sending interrupts for virtual devices. The other is the Guest
615 * timer interrupt.
616 *
617 * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
618 * the next timer interrupt (in nanoseconds). We use the high-resolution timer
619 * infrastructure to set a callback at that time.
620 *
621 * 0 means "turn off the clock".
622 */
623void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta)
624{
625 ktime_t expires;
626
627 if (unlikely(delta == 0)) {
628 /* Clock event device is shutting down. */
629 hrtimer_cancel(&cpu->hrt);
630 return;
631 }
632
633 /*
634 * We use wallclock time here, so the Guest might not be running for
635 * all the time between now and the timer interrupt it asked for. This
636 * is almost always the right thing to do.
637 */
638 expires = ktime_add_ns(ktime_get_real(), delta);
639 hrtimer_start(&cpu->hrt, expires, HRTIMER_MODE_ABS);
640}
641
642/* This is the function called when the Guest's timer expires. */
643static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
644{
645 struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt);
646
647 /* Remember the first interrupt is the timer interrupt. */
648 set_interrupt(cpu, 0);
649 return HRTIMER_NORESTART;
650}
651
652/* This sets up the timer for this Guest. */
653void init_clockdev(struct lg_cpu *cpu)
654{
655 hrtimer_init(&cpu->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
656 cpu->hrt.function = clockdev_fn;
657}