1 /*P:800 Interrupts (traps) are complicated enough to earn their own file.
2 * There are three classes of interrupts:
4 * 1) Real hardware interrupts which occur while we're running the Guest,
5 * 2) Interrupts for virtual devices attached to the Guest, and
6 * 3) Traps and faults from the Guest.
8 * Real hardware interrupts must be delivered to the Host, not the Guest.
9 * Virtual interrupts must be delivered to the Guest, but we make them look
10 * just like real hardware would deliver them. Traps from the Guest can be set
11 * up to go directly back into the Guest, but sometimes the Host wants to see
12 * them first, so we also have a way of "reflecting" them into the Guest as if
13 * they had been delivered to it directly. :*/
14 #include <linux/uaccess.h>
15 #include <linux/interrupt.h>
16 #include <linux/module.h>
19 /* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */
20 static unsigned int syscall_vector = SYSCALL_VECTOR;
21 module_param(syscall_vector, uint, 0444);
23 /* The address of the interrupt handler is split into two bits: */
24 static unsigned long idt_address(u32 lo, u32 hi)
26 return (lo & 0x0000FFFF) | (hi & 0xFFFF0000);
29 /* The "type" of the interrupt handler is a 4 bit field: we only support a
31 static int idt_type(u32 lo, u32 hi)
33 return (hi >> 8) & 0xF;
36 /* An IDT entry can't be used unless the "present" bit is set. */
37 static int idt_present(u32 lo, u32 hi)
42 /* We need a helper to "push" a value onto the Guest's stack, since that's a
43 * big part of what delivering an interrupt does. */
44 static void push_guest_stack(struct lguest *lg, unsigned long *gstack, u32 val)
46 /* Stack grows upwards: move stack then write value. */
48 lgwrite_u32(lg, *gstack, val);
51 /*H:210 The set_guest_interrupt() routine actually delivers the interrupt or
52 * trap. The mechanics of delivering traps and interrupts to the Guest are the
53 * same, except some traps have an "error code" which gets pushed onto the
54 * stack as well: the caller tells us if this is one.
56 * "lo" and "hi" are the two parts of the Interrupt Descriptor Table for this
57 * interrupt or trap. It's split into two parts for traditional reasons: gcc
58 * on i386 used to be frightened by 64 bit numbers.
60 * We set up the stack just like the CPU does for a real interrupt, so it's
61 * identical for the Guest (and the standard "iret" instruction will undo
63 static void set_guest_interrupt(struct lguest *lg, u32 lo, u32 hi, int has_err)
66 u32 eflags, ss, irq_enable;
68 /* There are two cases for interrupts: one where the Guest is already
69 * in the kernel, and a more complex one where the Guest is in
70 * userspace. We check the privilege level to find out. */
71 if ((lg->regs->ss&0x3) != GUEST_PL) {
72 /* The Guest told us their kernel stack with the SET_STACK
73 * hypercall: both the virtual address and the segment */
74 gstack = guest_pa(lg, lg->esp1);
76 /* We push the old stack segment and pointer onto the new
77 * stack: when the Guest does an "iret" back from the interrupt
78 * handler the CPU will notice they're dropping privilege
79 * levels and expect these here. */
80 push_guest_stack(lg, &gstack, lg->regs->ss);
81 push_guest_stack(lg, &gstack, lg->regs->esp);
83 /* We're staying on the same Guest (kernel) stack. */
84 gstack = guest_pa(lg, lg->regs->esp);
88 /* Remember that we never let the Guest actually disable interrupts, so
89 * the "Interrupt Flag" bit is always set. We copy that bit from the
90 * Guest's "irq_enabled" field into the eflags word: the Guest copies
91 * it back in "lguest_iret". */
92 eflags = lg->regs->eflags;
93 if (get_user(irq_enable, &lg->lguest_data->irq_enabled) == 0
94 && !(irq_enable & X86_EFLAGS_IF))
95 eflags &= ~X86_EFLAGS_IF;
97 /* An interrupt is expected to push three things on the stack: the old
98 * "eflags" word, the old code segment, and the old instruction
100 push_guest_stack(lg, &gstack, eflags);
101 push_guest_stack(lg, &gstack, lg->regs->cs);
102 push_guest_stack(lg, &gstack, lg->regs->eip);
104 /* For the six traps which supply an error code, we push that, too. */
106 push_guest_stack(lg, &gstack, lg->regs->errcode);
108 /* Now we've pushed all the old state, we change the stack, the code
109 * segment and the address to execute. */
111 lg->regs->esp = gstack + lg->page_offset;
112 lg->regs->cs = (__KERNEL_CS|GUEST_PL);
113 lg->regs->eip = idt_address(lo, hi);
115 /* There are two kinds of interrupt handlers: 0xE is an "interrupt
116 * gate" which expects interrupts to be disabled on entry. */
117 if (idt_type(lo, hi) == 0xE)
118 if (put_user(0, &lg->lguest_data->irq_enabled))
119 kill_guest(lg, "Disabling interrupts");
123 * Virtual Interrupts.
125 * maybe_do_interrupt() gets called before every entry to the Guest, to see if
126 * we should divert the Guest to running an interrupt handler. */
127 void maybe_do_interrupt(struct lguest *lg)
130 DECLARE_BITMAP(blk, LGUEST_IRQS);
131 struct desc_struct *idt;
133 /* If the Guest hasn't even initialized yet, we can do nothing. */
134 if (!lg->lguest_data)
137 /* Take our "irqs_pending" array and remove any interrupts the Guest
138 * wants blocked: the result ends up in "blk". */
139 if (copy_from_user(&blk, lg->lguest_data->blocked_interrupts,
143 bitmap_andnot(blk, lg->irqs_pending, blk, LGUEST_IRQS);
145 /* Find the first interrupt. */
146 irq = find_first_bit(blk, LGUEST_IRQS);
147 /* None? Nothing to do */
148 if (irq >= LGUEST_IRQS)
151 /* They may be in the middle of an iret, where they asked us never to
152 * deliver interrupts. */
153 if (lg->regs->eip >= lg->noirq_start && lg->regs->eip < lg->noirq_end)
156 /* If they're halted, interrupts restart them. */
158 /* Re-enable interrupts. */
159 if (put_user(X86_EFLAGS_IF, &lg->lguest_data->irq_enabled))
160 kill_guest(lg, "Re-enabling interrupts");
163 /* Otherwise we check if they have interrupts disabled. */
165 if (get_user(irq_enabled, &lg->lguest_data->irq_enabled))
171 /* Look at the IDT entry the Guest gave us for this interrupt. The
172 * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip
174 idt = &lg->arch.idt[FIRST_EXTERNAL_VECTOR+irq];
175 /* If they don't have a handler (yet?), we just ignore it */
176 if (idt_present(idt->a, idt->b)) {
177 /* OK, mark it no longer pending and deliver it. */
178 clear_bit(irq, lg->irqs_pending);
179 /* set_guest_interrupt() takes the interrupt descriptor and a
180 * flag to say whether this interrupt pushes an error code onto
181 * the stack as well: virtual interrupts never do. */
182 set_guest_interrupt(lg, idt->a, idt->b, 0);
185 /* Every time we deliver an interrupt, we update the timestamp in the
186 * Guest's lguest_data struct. It would be better for the Guest if we
187 * did this more often, but it can actually be quite slow: doing it
188 * here is a compromise which means at least it gets updated every
189 * timer interrupt. */
194 /* Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent
195 * me a patch, so we support that too. It'd be a big step for lguest if half
196 * the Plan 9 user base were to start using it.
198 * Actually now I think of it, it's possible that Ron *is* half the Plan 9
199 * userbase. Oh well. */
200 static bool could_be_syscall(unsigned int num)
202 /* Normal Linux SYSCALL_VECTOR or reserved vector? */
203 return num == SYSCALL_VECTOR || num == syscall_vector;
206 /* The syscall vector it wants must be unused by Host. */
207 bool check_syscall_vector(struct lguest *lg)
211 if (get_user(vector, &lg->lguest_data->syscall_vec))
214 return could_be_syscall(vector);
217 int init_interrupts(void)
219 /* If they want some strange system call vector, reserve it now */
220 if (syscall_vector != SYSCALL_VECTOR
221 && test_and_set_bit(syscall_vector, used_vectors)) {
222 printk("lg: couldn't reserve syscall %u\n", syscall_vector);
228 void free_interrupts(void)
230 if (syscall_vector != SYSCALL_VECTOR)
231 clear_bit(syscall_vector, used_vectors);
234 /*H:220 Now we've got the routines to deliver interrupts, delivering traps
235 * like page fault is easy. The only trick is that Intel decided that some
236 * traps should have error codes: */
237 static int has_err(unsigned int trap)
239 return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17);
242 /* deliver_trap() returns true if it could deliver the trap. */
243 int deliver_trap(struct lguest *lg, unsigned int num)
245 /* Trap numbers are always 8 bit, but we set an impossible trap number
246 * for traps inside the Switcher, so check that here. */
247 if (num >= ARRAY_SIZE(lg->arch.idt))
250 /* Early on the Guest hasn't set the IDT entries (or maybe it put a
251 * bogus one in): if we fail here, the Guest will be killed. */
252 if (!idt_present(lg->arch.idt[num].a, lg->arch.idt[num].b))
254 set_guest_interrupt(lg, lg->arch.idt[num].a, lg->arch.idt[num].b, has_err(num));
258 /*H:250 Here's the hard part: returning to the Host every time a trap happens
259 * and then calling deliver_trap() and re-entering the Guest is slow.
260 * Particularly because Guest userspace system calls are traps (trap 128).
262 * So we'd like to set up the IDT to tell the CPU to deliver traps directly
263 * into the Guest. This is possible, but the complexities cause the size of
264 * this file to double! However, 150 lines of code is worth writing for taking
265 * system calls down from 1750ns to 270ns. Plus, if lguest didn't do it, all
266 * the other hypervisors would tease it.
268 * This routine indicates if a particular trap number could be delivered
270 static int direct_trap(unsigned int num)
272 /* Hardware interrupts don't go to the Guest at all (except system
274 if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num))
277 /* The Host needs to see page faults (for shadow paging and to save the
278 * fault address), general protection faults (in/out emulation) and
279 * device not available (TS handling), and of course, the hypercall
281 return num != 14 && num != 13 && num != 7 && num != LGUEST_TRAP_ENTRY;
285 /*M:005 The Guest has the ability to turn its interrupt gates into trap gates,
286 * if it is careful. The Host will let trap gates can go directly to the
287 * Guest, but the Guest needs the interrupts atomically disabled for an
288 * interrupt gate. It can do this by pointing the trap gate at instructions
289 * within noirq_start and noirq_end, where it can safely disable interrupts. */
291 /*M:006 The Guests do not use the sysenter (fast system call) instruction,
292 * because it's hardcoded to enter privilege level 0 and so can't go direct.
293 * It's about twice as fast as the older "int 0x80" system call, so it might
294 * still be worthwhile to handle it in the Switcher and lcall down to the
295 * Guest. The sysenter semantics are hairy tho: search for that keyword in
298 /*H:260 When we make traps go directly into the Guest, we need to make sure
299 * the kernel stack is valid (ie. mapped in the page tables). Otherwise, the
300 * CPU trying to deliver the trap will fault while trying to push the interrupt
301 * words on the stack: this is called a double fault, and it forces us to kill
304 * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. */
305 void pin_stack_pages(struct lguest *lg)
309 /* Depending on the CONFIG_4KSTACKS option, the Guest can have one or
310 * two pages of stack space. */
311 for (i = 0; i < lg->stack_pages; i++)
312 /* The stack grows *upwards*, so the address we're given is the
313 * start of the page after the kernel stack. Subtract one to
314 * get back onto the first stack page, and keep subtracting to
315 * get to the rest of the stack pages. */
316 pin_page(lg, lg->esp1 - 1 - i * PAGE_SIZE);
319 /* Direct traps also mean that we need to know whenever the Guest wants to use
320 * a different kernel stack, so we can change the IDT entries to use that
321 * stack. The IDT entries expect a virtual address, so unlike most addresses
322 * the Guest gives us, the "esp" (stack pointer) value here is virtual, not
325 * In Linux each process has its own kernel stack, so this happens a lot: we
326 * change stacks on each context switch. */
327 void guest_set_stack(struct lguest *lg, u32 seg, u32 esp, unsigned int pages)
329 /* You are not allowd have a stack segment with privilege level 0: bad
331 if ((seg & 0x3) != GUEST_PL)
332 kill_guest(lg, "bad stack segment %i", seg);
333 /* We only expect one or two stack pages. */
335 kill_guest(lg, "bad stack pages %u", pages);
336 /* Save where the stack is, and how many pages */
339 lg->stack_pages = pages;
340 /* Make sure the new stack pages are mapped */
344 /* All this reference to mapping stacks leads us neatly into the other complex
345 * part of the Host: page table handling. */
347 /*H:235 This is the routine which actually checks the Guest's IDT entry and
348 * transfers it into our entry in "struct lguest": */
349 static void set_trap(struct lguest *lg, struct desc_struct *trap,
350 unsigned int num, u32 lo, u32 hi)
352 u8 type = idt_type(lo, hi);
354 /* We zero-out a not-present entry */
355 if (!idt_present(lo, hi)) {
356 trap->a = trap->b = 0;
360 /* We only support interrupt and trap gates. */
361 if (type != 0xE && type != 0xF)
362 kill_guest(lg, "bad IDT type %i", type);
364 /* We only copy the handler address, present bit, privilege level and
365 * type. The privilege level controls where the trap can be triggered
366 * manually with an "int" instruction. This is usually GUEST_PL,
367 * except for system calls which userspace can use. */
368 trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF);
369 trap->b = (hi&0xFFFFEF00);
372 /*H:230 While we're here, dealing with delivering traps and interrupts to the
373 * Guest, we might as well complete the picture: how the Guest tells us where
374 * it wants them to go. This would be simple, except making traps fast
375 * requires some tricks.
377 * We saw the Guest setting Interrupt Descriptor Table (IDT) entries with the
378 * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. */
379 void load_guest_idt_entry(struct lguest *lg, unsigned int num, u32 lo, u32 hi)
381 /* Guest never handles: NMI, doublefault, spurious interrupt or
382 * hypercall. We ignore when it tries to set them. */
383 if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY)
386 /* Mark the IDT as changed: next time the Guest runs we'll know we have
387 * to copy this again. */
388 lg->changed |= CHANGED_IDT;
390 /* Check that the Guest doesn't try to step outside the bounds. */
391 if (num >= ARRAY_SIZE(lg->arch.idt))
392 kill_guest(lg, "Setting idt entry %u", num);
394 set_trap(lg, &lg->arch.idt[num], num, lo, hi);
397 /* The default entry for each interrupt points into the Switcher routines which
398 * simply return to the Host. The run_guest() loop will then call
399 * deliver_trap() to bounce it back into the Guest. */
400 static void default_idt_entry(struct desc_struct *idt,
402 const unsigned long handler)
404 /* A present interrupt gate. */
407 /* Set the privilege level on the entry for the hypercall: this allows
408 * the Guest to use the "int" instruction to trigger it. */
409 if (trap == LGUEST_TRAP_ENTRY)
410 flags |= (GUEST_PL << 13);
412 /* Now pack it into the IDT entry in its weird format. */
413 idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF);
414 idt->b = (handler&0xFFFF0000) | flags;
417 /* When the Guest first starts, we put default entries into the IDT. */
418 void setup_default_idt_entries(struct lguest_ro_state *state,
419 const unsigned long *def)
423 for (i = 0; i < ARRAY_SIZE(state->guest_idt); i++)
424 default_idt_entry(&state->guest_idt[i], i, def[i]);
427 /*H:240 We don't use the IDT entries in the "struct lguest" directly, instead
428 * we copy them into the IDT which we've set up for Guests on this CPU, just
429 * before we run the Guest. This routine does that copy. */
430 void copy_traps(const struct lguest *lg, struct desc_struct *idt,
431 const unsigned long *def)
435 /* We can simply copy the direct traps, otherwise we use the default
436 * ones in the Switcher: they will return to the Host. */
437 for (i = 0; i < ARRAY_SIZE(lg->arch.idt); i++) {
438 /* If no Guest can ever override this trap, leave it alone. */
442 /* Only trap gates (type 15) can go direct to the Guest.
443 * Interrupt gates (type 14) disable interrupts as they are
444 * entered, which we never let the Guest do. Not present
445 * entries (type 0x0) also can't go direct, of course. */
446 if (idt_type(lg->arch.idt[i].a, lg->arch.idt[i].b) == 0xF)
447 idt[i] = lg->arch.idt[i];
449 /* Reset it to the default. */
450 default_idt_entry(&idt[i], i, def[i]);
454 void guest_set_clockevent(struct lguest *lg, unsigned long delta)
458 if (unlikely(delta == 0)) {
459 /* Clock event device is shutting down. */
460 hrtimer_cancel(&lg->hrt);
464 expires = ktime_add_ns(ktime_get_real(), delta);
465 hrtimer_start(&lg->hrt, expires, HRTIMER_MODE_ABS);
468 static enum hrtimer_restart clockdev_fn(struct hrtimer *timer)
470 struct lguest *lg = container_of(timer, struct lguest, hrt);
472 set_bit(0, lg->irqs_pending);
474 wake_up_process(lg->tsk);
475 return HRTIMER_NORESTART;
478 void init_clockdev(struct lguest *lg)
480 hrtimer_init(&lg->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS);
481 lg->hrt.function = clockdev_fn;