1 /*P:800 Interrupts (traps) are complicated enough to earn their own file.
2 * There are three classes of interrupts:
3 *
4 * 1) Real hardware interrupts which occur while we're running the Guest,
5 * 2) Interrupts for virtual devices attached to the Guest, and
6 * 3) Traps and faults from the Guest.
7 *
8 * Real hardware interrupts must be delivered to the Host, not the Guest.
9 * Virtual interrupts must be delivered to the Guest, but we make them look
10 * just like real hardware would deliver them. Traps from the Guest can be set
11 * up to go directly back into the Guest, but sometimes the Host wants to see
12 * them first, so we also have a way of "reflecting" them into the Guest as if
13 * they had been delivered to it directly. :*/
14 #include <linux/uaccess.h>
15 #include <linux/interrupt.h>
16 #include <linux/module.h>
17 #include "lg.h"
18
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);
22
23 /* The address of the interrupt handler is split into two bits: */
idt_address(u32 lo,u32 hi)24 static unsigned long idt_address(u32 lo, u32 hi)
25 {
26 return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
27 }
28
29 /* The "type" of the interrupt handler is a 4 bit field: we only support a
30 * couple of types. */
idt_type(u32 lo,u32 hi)31 static int idt_type(u32 lo, u32 hi)
32 {
33 return (hi >> 8) & 0xF;
34 }
35
36 /* An IDT entry can't be used unless the "present" bit is set. */
idt_present(u32 lo,u32 hi)37 static int idt_present(u32 lo, u32 hi)
38 {
39 return (hi & 0x8000);
40 }
41
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. */
push_guest_stack(struct lg_cpu * cpu,unsigned long * gstack,u32 val)44 static void push_guest_stack(struct lg_cpu *cpu, unsigned long *gstack, u32 val)
45 {
46 /* Stack grows upwards: move stack then write value. */
47 *gstack -= 4;
48 lgwrite(cpu, *gstack, u32, val);
49 }
50
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.
55 *
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.
59 *
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
62 * it). */
set_guest_interrupt(struct lg_cpu * cpu,u32 lo,u32 hi,int has_err)63 static void set_guest_interrupt(struct lg_cpu *cpu, u32 lo, u32 hi, int has_err)
64 {
65 unsigned long gstack, origstack;
66 u32 eflags, ss, irq_enable;
67 unsigned long virtstack;
68
69 /* There are two cases for interrupts: one where the Guest is already
70 * in the kernel, and a more complex one where the Guest is in
71 * userspace. We check the privilege level to find out. */
72 if ((cpu->regs->ss&0x3) != GUEST_PL) {
73 /* The Guest told us their kernel stack with the SET_STACK
74 * hypercall: both the virtual address and the segment */
75 virtstack = cpu->esp1;
76 ss = cpu->ss1;
77
78 origstack = gstack = guest_pa(cpu, virtstack);
79 /* We push the old stack segment and pointer onto the new
80 * stack: when the Guest does an "iret" back from the interrupt
81 * handler the CPU will notice they're dropping privilege
82 * levels and expect these here. */
83 push_guest_stack(cpu, &gstack, cpu->regs->ss);
84 push_guest_stack(cpu, &gstack, cpu->regs->esp);
85 } else {
86 /* We're staying on the same Guest (kernel) stack. */
87 virtstack = cpu->regs->esp;
88 ss = cpu->regs->ss;
89
90 origstack = gstack = guest_pa(cpu, virtstack);
91 }
92
93 /* Remember that we never let the Guest actually disable interrupts, so
94 * the "Interrupt Flag" bit is always set. We copy that bit from the
95 * Guest's "irq_enabled" field into the eflags word: we saw the Guest
96 * copy it back in "lguest_iret". */
97 eflags = cpu->regs->eflags;
98 if (get_user(irq_enable, &cpu->lg->lguest_data->irq_enabled) == 0
99 && !(irq_enable & X86_EFLAGS_IF))
100 eflags &= ~X86_EFLAGS_IF;
101
102 /* An interrupt is expected to push three things on the stack: the old
103 * "eflags" word, the old code segment, and the old instruction
104 * pointer. */
105 push_guest_stack(cpu, &gstack, eflags);
106 push_guest_stack(cpu, &gstack, cpu->regs->cs);
107 push_guest_stack(cpu, &gstack, cpu->regs->eip);
108
109 /* For the six traps which supply an error code, we push that, too. */
110 if (has_err)
111 push_guest_stack(cpu, &gstack, cpu->regs->errcode);
112
113 /* Now we've pushed all the old state, we change the stack, the code
114 * segment and the address to execute. */
115 cpu->regs->ss = ss;
116 cpu->regs->esp = virtstack + (gstack - origstack);
117 cpu->regs->cs = (__KERNEL_CS|GUEST_PL);
118 cpu->regs->eip = idt_address(lo, hi);
119
120 /* There are two kinds of interrupt handlers: 0xE is an "interrupt
121 * gate" which expects interrupts to be disabled on entry. */
122 if (idt_type(lo, hi) == 0xE)
123 if (put_user(0, &cpu->lg->lguest_data->irq_enabled))
124 kill_guest(cpu, "Disabling interrupts");
125 }
126
127 /*H:205
128 * Virtual Interrupts.
129 *
130 * maybe_do_interrupt() gets called before every entry to the Guest, to see if
131 * we should divert the Guest to running an interrupt handler. */
maybe_do_interrupt(struct lg_cpu * cpu)132 void maybe_do_interrupt(struct lg_cpu *cpu)
133 {
134 unsigned int irq;
135 DECLARE_BITMAP(blk, LGUEST_IRQS);
136 struct desc_struct *idt;
137
138 /* If the Guest hasn't even initialized yet, we can do nothing. */
139 if (!cpu->lg->lguest_data)
140 return;
141
142 /* Take our "irqs_pending" array and remove any interrupts the Guest
143 * wants blocked: the result ends up in "blk". */
144 if (copy_from_user(&blk, cpu->lg->lguest_data->blocked_interrupts,
145 sizeof(blk)))
146 return;
147 bitmap_andnot(blk, cpu->irqs_pending, blk, LGUEST_IRQS);
148
149 /* Find the first interrupt. */
150 irq = find_first_bit(blk, LGUEST_IRQS);
151 /* None? Nothing to do */
152 if (irq >= LGUEST_IRQS)
153 return;
154
155 /* They may be in the middle of an iret, where they asked us never to
156 * deliver interrupts. */
157 if (cpu->regs->eip >= cpu->lg->noirq_start &&
158 (cpu->regs->eip < cpu->lg->noirq_end))
159 return;
160
161 /* If they're halted, interrupts restart them. */
162 if (cpu->halted) {
163 /* Re-enable interrupts. */
164 if (put_user(X86_EFLAGS_IF, &cpu->lg->lguest_data->irq_enabled))
165 kill_guest(cpu, "Re-enabling interrupts");
166 cpu->halted = 0;
167 } else {
168 /* Otherwise we check if they have interrupts disabled. */
169 u32 irq_enabled;
170 if (get_user(irq_enabled, &cpu->lg->lguest_data->irq_enabled))
171 irq_enabled = 0;
172 if (!irq_enabled)
173 return;
174 }
175
176 /* Look at the IDT entry the Guest gave us for this interrupt. The
177 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
178 * over them. */
179 idt = &cpu->arch.idt[FIRST_EXTERNAL_VECTOR+irq];
180 /* If they don't have a handler (yet?), we just ignore it */
181 if (idt_present(idt->a, idt->b)) {
182 /* OK, mark it no longer pending and deliver it. */
183 clear_bit(irq, cpu->irqs_pending);
184 /* set_guest_interrupt() takes the interrupt descriptor and a
185 * flag to say whether this interrupt pushes an error code onto
186 * the stack as well: virtual interrupts never do. */
187 set_guest_interrupt(cpu, idt->a, idt->b, 0);
188 }
189
190 /* Every time we deliver an interrupt, we update the timestamp in the
191 * Guest's lguest_data struct. It would be better for the Guest if we
192 * did this more often, but it can actually be quite slow: doing it
193 * here is a compromise which means at least it gets updated every
194 * timer interrupt. */
195 write_timestamp(cpu);
196 }
197 /*:*/
198
199 /* Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent
200 * me a patch, so we support that too. It'd be a big step for lguest if half
201 * the Plan 9 user base were to start using it.
202 *
203 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
204 * userbase. Oh well. */
could_be_syscall(unsigned int num)205 static bool could_be_syscall(unsigned int num)
206 {
207 /* Normal Linux SYSCALL_VECTOR or reserved vector? */
208 return num == SYSCALL_VECTOR || num == syscall_vector;
209 }
210
211 /* The syscall vector it wants must be unused by Host. */
check_syscall_vector(struct lguest * lg)212 bool check_syscall_vector(struct lguest *lg)
213 {
214 u32 vector;
215
216 if (get_user(vector, &lg->lguest_data->syscall_vec))
217 return false;
218
219 return could_be_syscall(vector);
220 }
221
init_interrupts(void)222 int init_interrupts(void)
223 {
224 /* If they want some strange system call vector, reserve it now */
225 if (syscall_vector != SYSCALL_VECTOR) {
226 if (test_bit(syscall_vector, used_vectors) ||
227 vector_used_by_percpu_irq(syscall_vector)) {
228 printk(KERN_ERR "lg: couldn't reserve syscall %u\n",
229 syscall_vector);
230 return -EBUSY;
231 }
232 set_bit(syscall_vector, used_vectors);
233 }
234
235 return 0;
236 }
237
free_interrupts(void)238 void free_interrupts(void)
239 {
240 if (syscall_vector != SYSCALL_VECTOR)
241 clear_bit(syscall_vector, used_vectors);
242 }
243
244 /*H:220 Now we've got the routines to deliver interrupts, delivering traps like
245 * page fault is easy. The only trick is that Intel decided that some traps
246 * should have error codes: */
has_err(unsigned int trap)247 static int has_err(unsigned int trap)
248 {
249 return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
250 }
251
252 /* deliver_trap() returns true if it could deliver the trap. */
deliver_trap(struct lg_cpu * cpu,unsigned int num)253 int deliver_trap(struct lg_cpu *cpu, unsigned int num)
254 {
255 /* Trap numbers are always 8 bit, but we set an impossible trap number
256 * for traps inside the Switcher, so check that here. */
257 if (num >= ARRAY_SIZE(cpu->arch.idt))
258 return 0;
259
260 /* Early on the Guest hasn't set the IDT entries (or maybe it put a
261 * bogus one in): if we fail here, the Guest will be killed. */
262 if (!idt_present(cpu->arch.idt[num].a, cpu->arch.idt[num].b))
263 return 0;
264 set_guest_interrupt(cpu, cpu->arch.idt[num].a,
265 cpu->arch.idt[num].b, has_err(num));
266 return 1;
267 }
268
269 /*H:250 Here's the hard part: returning to the Host every time a trap happens
270 * and then calling deliver_trap() and re-entering the Guest is slow.
271 * Particularly because Guest userspace system calls are traps (usually trap
272 * 128).
273 *
274 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
275 * into the Guest. This is possible, but the complexities cause the size of
276 * this file to double! However, 150 lines of code is worth writing for taking
277 * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
278 * the other hypervisors would beat it up at lunchtime.
279 *
280 * This routine indicates if a particular trap number could be delivered
281 * directly. */
direct_trap(unsigned int num)282 static int direct_trap(unsigned int num)
283 {
284 /* Hardware interrupts don't go to the Guest at all (except system
285 * call). */
286 if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
287 return 0;
288
289 /* The Host needs to see page faults (for shadow paging and to save the
290 * fault address), general protection faults (in/out emulation) and
291 * device not available (TS handling), and of course, the hypercall
292 * trap. */
293 return num != 14 && num != 13 && num != 7 && num != LGUEST_TRAP_ENTRY;
294 }
295 /*:*/
296
297 /*M:005 The Guest has the ability to turn its interrupt gates into trap gates,
298 * if it is careful. The Host will let trap gates can go directly to the
299 * Guest, but the Guest needs the interrupts atomically disabled for an
300 * interrupt gate. It can do this by pointing the trap gate at instructions
301 * within noirq_start and noirq_end, where it can safely disable interrupts. */
302
303 /*M:006 The Guests do not use the sysenter (fast system call) instruction,
304 * because it's hardcoded to enter privilege level 0 and so can't go direct.
305 * It's about twice as fast as the older "int 0x80" system call, so it might
306 * still be worthwhile to handle it in the Switcher and lcall down to the
307 * Guest. The sysenter semantics are hairy tho: search for that keyword in
308 * entry.S :*/
309
310 /*H:260 When we make traps go directly into the Guest, we need to make sure
311 * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the
312 * CPU trying to deliver the trap will fault while trying to push the interrupt
313 * words on the stack: this is called a double fault, and it forces us to kill
314 * the Guest.
315 *
316 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */
pin_stack_pages(struct lg_cpu * cpu)317 void pin_stack_pages(struct lg_cpu *cpu)
318 {
319 unsigned int i;
320
321 /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or
322 * two pages of stack space. */
323 for (i = 0; i < cpu->lg->stack_pages; i++)
324 /* The stack grows *upwards*, so the address we're given is the
325 * start of the page after the kernel stack. Subtract one to
326 * get back onto the first stack page, and keep subtracting to
327 * get to the rest of the stack pages. */
328 pin_page(cpu, cpu->esp1 - 1 - i * PAGE_SIZE);
329 }
330
331 /* Direct traps also mean that we need to know whenever the Guest wants to use
332 * a different kernel stack, so we can change the IDT entries to use that
333 * stack. The IDT entries expect a virtual address, so unlike most addresses
334 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
335 * physical.
336 *
337 * In Linux each process has its own kernel stack, so this happens a lot: we
338 * change stacks on each context switch. */
guest_set_stack(struct lg_cpu * cpu,u32 seg,u32 esp,unsigned int pages)339 void guest_set_stack(struct lg_cpu *cpu, u32 seg, u32 esp, unsigned int pages)
340 {
341 /* You are not allowed have a stack segment with privilege level 0: bad
342 * Guest! */
343 if ((seg & 0x3) != GUEST_PL)
344 kill_guest(cpu, "bad stack segment %i", seg);
345 /* We only expect one or two stack pages. */
346 if (pages > 2)
347 kill_guest(cpu, "bad stack pages %u", pages);
348 /* Save where the stack is, and how many pages */
349 cpu->ss1 = seg;
350 cpu->esp1 = esp;
351 cpu->lg->stack_pages = pages;
352 /* Make sure the new stack pages are mapped */
353 pin_stack_pages(cpu);
354 }
355
356 /* All this reference to mapping stacks leads us neatly into the other complex
357 * part of the Host: page table handling. */
358
359 /*H:235 This is the routine which actually checks the Guest's IDT entry and
360 * transfers it into the entry in "struct lguest": */
set_trap(struct lg_cpu * cpu,struct desc_struct * trap,unsigned int num,u32 lo,u32 hi)361 static void set_trap(struct lg_cpu *cpu, struct desc_struct *trap,
362 unsigned int num, u32 lo, u32 hi)
363 {
364 u8 type = idt_type(lo, hi);
365
366 /* We zero-out a not-present entry */
367 if (!idt_present(lo, hi)) {
368 trap->a = trap->b = 0;
369 return;
370 }
371
372 /* We only support interrupt and trap gates. */
373 if (type != 0xE && type != 0xF)
374 kill_guest(cpu, "bad IDT type %i", type);
375
376 /* We only copy the handler address, present bit, privilege level and
377 * type. The privilege level controls where the trap can be triggered
378 * manually with an "int" instruction. This is usually GUEST_PL,
379 * except for system calls which userspace can use. */
380 trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
381 trap->b = (hi&0xFFFFEF00);
382 }
383
384 /*H:230 While we're here, dealing with delivering traps and interrupts to the
385 * Guest, we might as well complete the picture: how the Guest tells us where
386 * it wants them to go. This would be simple, except making traps fast
387 * requires some tricks.
388 *
389 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
390 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */
load_guest_idt_entry(struct lg_cpu * cpu,unsigned int num,u32 lo,u32 hi)391 void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int num, u32 lo, u32 hi)
392 {
393 /* Guest never handles: NMI, doublefault, spurious interrupt or
394 * hypercall. We ignore when it tries to set them. */
395 if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
396 return;
397
398 /* Mark the IDT as changed: next time the Guest runs we'll know we have
399 * to copy this again. */
400 cpu->changed |= CHANGED_IDT;
401
402 /* Check that the Guest doesn't try to step outside the bounds. */
403 if (num >= ARRAY_SIZE(cpu->arch.idt))
404 kill_guest(cpu, "Setting idt entry %u", num);
405 else
406 set_trap(cpu, &cpu->arch.idt[num], num, lo, hi);
407 }
408
409 /* The default entry for each interrupt points into the Switcher routines which
410 * simply return to the Host. The run_guest() loop will then call
411 * deliver_trap() to bounce it back into the Guest. */
default_idt_entry(struct desc_struct * idt,int trap,const unsigned long handler,const struct desc_struct * base)412 static void default_idt_entry(struct desc_struct *idt,
413 int trap,
414 const unsigned long handler,
415 const struct desc_struct *base)
416 {
417 /* A present interrupt gate. */
418 u32 flags = 0x8e00;
419
420 /* Set the privilege level on the entry for the hypercall: this allows
421 * the Guest to use the "int" instruction to trigger it. */
422 if (trap == LGUEST_TRAP_ENTRY)
423 flags |= (GUEST_PL << 13);
424 else if (base)
425 /* Copy priv. level from what Guest asked for. This allows
426 * debug (int 3) traps from Guest userspace, for example. */
427 flags |= (base->b & 0x6000);
428
429 /* Now pack it into the IDT entry in its weird format. */
430 idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
431 idt->b = (handler&0xFFFF0000) | flags;
432 }
433
434 /* When the Guest first starts, we put default entries into the IDT. */
setup_default_idt_entries(struct lguest_ro_state * state,const unsigned long * def)435 void setup_default_idt_entries(struct lguest_ro_state *state,
436 const unsigned long *def)
437 {
438 unsigned int i;
439
440 for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
441 default_idt_entry(&state->guest_idt[i], i, def[i], NULL);
442 }
443
444 /*H:240 We don't use the IDT entries in the "struct lguest" directly, instead
445 * we copy them into the IDT which we've set up for Guests on this CPU, just
446 * before we run the Guest. This routine does that copy. */
copy_traps(const struct lg_cpu * cpu,struct desc_struct * idt,const unsigned long * def)447 void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt,
448 const unsigned long *def)
449 {
450 unsigned int i;
451
452 /* We can simply copy the direct traps, otherwise we use the default
453 * ones in the Switcher: they will return to the Host. */
454 for (i = 0; i < ARRAY_SIZE(cpu->arch.idt); i++) {
455 const struct desc_struct *gidt = &cpu->arch.idt[i];
456
457 /* If no Guest can ever override this trap, leave it alone. */
458 if (!direct_trap(i))
459 continue;
460
461 /* Only trap gates (type 15) can go direct to the Guest.
462 * Interrupt gates (type 14) disable interrupts as they are
463 * entered, which we never let the Guest do. Not present
464 * entries (type 0x0) also can't go direct, of course.
465 *
466 * If it can't go direct, we still need to copy the priv. level:
467 * they might want to give userspace access to a software
468 * interrupt. */
469 if (idt_type(gidt->a, gidt->b) == 0xF)
470 idt[i] = *gidt;
471 else
472 default_idt_entry(&idt[i], i, def[i], gidt);
473 }
474 }
475
476 /*H:200
477 * The Guest Clock.
478 *
479 * There are two sources of virtual interrupts. We saw one in lguest_user.c:
480 * the Launcher sending interrupts for virtual devices. The other is the Guest
481 * timer interrupt.
482 *
483 * The Guest uses the LHCALL_SET_CLOCKEVENT hypercall to tell us how long to
484 * the next timer interrupt (in nanoseconds). We use the high-resolution timer
485 * infrastructure to set a callback at that time.
486 *
487 * 0 means "turn off the clock". */
guest_set_clockevent(struct lg_cpu * cpu,unsigned long delta)488 void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta)
489 {
490 ktime_t expires;
491
492 if (unlikely(delta == 0)) {
493 /* Clock event device is shutting down. */
494 hrtimer_cancel(&cpu->hrt);
495 return;
496 }
497
498 /* We use wallclock time here, so the Guest might not be running for
499 * all the time between now and the timer interrupt it asked for. This
500 * is almost always the right thing to do. */
501 expires = ktime_add_ns(ktime_get_real(), delta);
502 hrtimer_start(&cpu->hrt, expires, HRTIMER_MODE_ABS);
503 }
504
505 /* This is the function called when the Guest's timer expires. */
clockdev_fn(struct hrtimer * timer)506 static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
507 {
508 struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt);
509
510 /* Remember the first interrupt is the timer interrupt. */
511 set_bit(0, cpu->irqs_pending);
512 /* If the Guest is actually stopped, we need to wake it up. */
513 if (cpu->halted)
514 wake_up_process(cpu->tsk);
515 return HRTIMER_NORESTART;
516 }
517
518 /* This sets up the timer for this Guest. */
init_clockdev(struct lg_cpu * cpu)519 void init_clockdev(struct lg_cpu *cpu)
520 {
521 hrtimer_init(&cpu->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
522 cpu->hrt.function = clockdev_fn;
523 }
524