Commit | Line | Data |
---|---|---|
2e04ef76 RR |
1 | /*P:800 |
2 | * Interrupts (traps) are complicated enough to earn their own file. | |
f938d2c8 RR |
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 | |
2e04ef76 RR |
14 | * they had been delivered to it directly. |
15 | :*/ | |
d7e28ffe | 16 | #include <linux/uaccess.h> |
c18acd73 RR |
17 | #include <linux/interrupt.h> |
18 | #include <linux/module.h> | |
d43c36dc | 19 | #include <linux/sched.h> |
d7e28ffe RR |
20 | #include "lg.h" |
21 | ||
c18acd73 RR |
22 | /* Allow Guests to use a non-128 (ie. non-Linux) syscall trap. */ |
23 | static unsigned int syscall_vector = SYSCALL_VECTOR; | |
24 | module_param(syscall_vector, uint, 0444); | |
25 | ||
bff672e6 | 26 | /* The address of the interrupt handler is split into two bits: */ |
d7e28ffe RR |
27 | static unsigned long idt_address(u32 lo, u32 hi) |
28 | { | |
29 | return (lo & 0x0000FFFF) | (hi & 0xFFFF0000); | |
30 | } | |
31 | ||
2e04ef76 RR |
32 | /* |
33 | * The "type" of the interrupt handler is a 4 bit field: we only support a | |
34 | * couple of types. | |
35 | */ | |
d7e28ffe RR |
36 | static int idt_type(u32 lo, u32 hi) |
37 | { | |
38 | return (hi >> 8) & 0xF; | |
39 | } | |
40 | ||
bff672e6 | 41 | /* An IDT entry can't be used unless the "present" bit is set. */ |
df1693ab | 42 | static bool idt_present(u32 lo, u32 hi) |
d7e28ffe RR |
43 | { |
44 | return (hi & 0x8000); | |
45 | } | |
46 | ||
2e04ef76 RR |
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 | */ | |
382ac6b3 | 51 | static void push_guest_stack(struct lg_cpu *cpu, unsigned long *gstack, u32 val) |
d7e28ffe | 52 | { |
bff672e6 | 53 | /* Stack grows upwards: move stack then write value. */ |
d7e28ffe | 54 | *gstack -= 4; |
382ac6b3 | 55 | lgwrite(cpu, *gstack, u32, val); |
d7e28ffe RR |
56 | } |
57 | ||
2e04ef76 RR |
58 | /*H:210 |
59 | * The set_guest_interrupt() routine actually delivers the interrupt or | |
bff672e6 RR |
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 | |
2e04ef76 RR |
70 | * it). |
71 | */ | |
df1693ab MZ |
72 | static void set_guest_interrupt(struct lg_cpu *cpu, u32 lo, u32 hi, |
73 | bool has_err) | |
d7e28ffe | 74 | { |
47436aa4 | 75 | unsigned long gstack, origstack; |
d7e28ffe | 76 | u32 eflags, ss, irq_enable; |
47436aa4 | 77 | unsigned long virtstack; |
d7e28ffe | 78 | |
2e04ef76 RR |
79 | /* |
80 | * There are two cases for interrupts: one where the Guest is already | |
bff672e6 | 81 | * in the kernel, and a more complex one where the Guest is in |
2e04ef76 RR |
82 | * userspace. We check the privilege level to find out. |
83 | */ | |
a53a35a8 | 84 | if ((cpu->regs->ss&0x3) != GUEST_PL) { |
2e04ef76 RR |
85 | /* |
86 | * The Guest told us their kernel stack with the SET_STACK | |
87 | * hypercall: both the virtual address and the segment. | |
88 | */ | |
4665ac8e GOC |
89 | virtstack = cpu->esp1; |
90 | ss = cpu->ss1; | |
47436aa4 | 91 | |
1713608f | 92 | origstack = gstack = guest_pa(cpu, virtstack); |
2e04ef76 RR |
93 | /* |
94 | * We push the old stack segment and pointer onto the new | |
bff672e6 RR |
95 | * stack: when the Guest does an "iret" back from the interrupt |
96 | * handler the CPU will notice they're dropping privilege | |
2e04ef76 RR |
97 | * levels and expect these here. |
98 | */ | |
382ac6b3 GOC |
99 | push_guest_stack(cpu, &gstack, cpu->regs->ss); |
100 | push_guest_stack(cpu, &gstack, cpu->regs->esp); | |
d7e28ffe | 101 | } else { |
bff672e6 | 102 | /* We're staying on the same Guest (kernel) stack. */ |
a53a35a8 GOC |
103 | virtstack = cpu->regs->esp; |
104 | ss = cpu->regs->ss; | |
47436aa4 | 105 | |
1713608f | 106 | origstack = gstack = guest_pa(cpu, virtstack); |
d7e28ffe RR |
107 | } |
108 | ||
2e04ef76 RR |
109 | /* |
110 | * Remember that we never let the Guest actually disable interrupts, so | |
bff672e6 | 111 | * the "Interrupt Flag" bit is always set. We copy that bit from the |
e1e72965 | 112 | * Guest's "irq_enabled" field into the eflags word: we saw the Guest |
2e04ef76 RR |
113 | * copy it back in "lguest_iret". |
114 | */ | |
a53a35a8 | 115 | eflags = cpu->regs->eflags; |
382ac6b3 | 116 | if (get_user(irq_enable, &cpu->lg->lguest_data->irq_enabled) == 0 |
e5faff45 RR |
117 | && !(irq_enable & X86_EFLAGS_IF)) |
118 | eflags &= ~X86_EFLAGS_IF; | |
d7e28ffe | 119 | |
2e04ef76 RR |
120 | /* |
121 | * An interrupt is expected to push three things on the stack: the old | |
bff672e6 | 122 | * "eflags" word, the old code segment, and the old instruction |
2e04ef76 RR |
123 | * pointer. |
124 | */ | |
382ac6b3 GOC |
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); | |
d7e28ffe | 128 | |
bff672e6 | 129 | /* For the six traps which supply an error code, we push that, too. */ |
d7e28ffe | 130 | if (has_err) |
382ac6b3 | 131 | push_guest_stack(cpu, &gstack, cpu->regs->errcode); |
d7e28ffe | 132 | |
2e04ef76 RR |
133 | /* |
134 | * Now we've pushed all the old state, we change the stack, the code | |
135 | * segment and the address to execute. | |
136 | */ | |
a53a35a8 GOC |
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); | |
d7e28ffe | 141 | |
98fb4e5e RR |
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 | ||
2e04ef76 RR |
152 | /* |
153 | * There are two kinds of interrupt handlers: 0xE is an "interrupt | |
154 | * gate" which expects interrupts to be disabled on entry. | |
155 | */ | |
d7e28ffe | 156 | if (idt_type(lo, hi) == 0xE) |
382ac6b3 GOC |
157 | if (put_user(0, &cpu->lg->lguest_data->irq_enabled)) |
158 | kill_guest(cpu, "Disabling interrupts"); | |
d7e28ffe RR |
159 | } |
160 | ||
e1e72965 | 161 | /*H:205 |
bff672e6 RR |
162 | * Virtual Interrupts. |
163 | * | |
abd41f03 RR |
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 | |
2e04ef76 RR |
166 | * we go to sleep when the Guest has halted itself. |
167 | */ | |
a32a8813 | 168 | unsigned int interrupt_pending(struct lg_cpu *cpu, bool *more) |
d7e28ffe RR |
169 | { |
170 | unsigned int irq; | |
171 | DECLARE_BITMAP(blk, LGUEST_IRQS); | |
d7e28ffe | 172 | |
bff672e6 | 173 | /* If the Guest hasn't even initialized yet, we can do nothing. */ |
382ac6b3 | 174 | if (!cpu->lg->lguest_data) |
abd41f03 | 175 | return LGUEST_IRQS; |
d7e28ffe | 176 | |
2e04ef76 RR |
177 | /* |
178 | * Take our "irqs_pending" array and remove any interrupts the Guest | |
179 | * wants blocked: the result ends up in "blk". | |
180 | */ | |
382ac6b3 | 181 | if (copy_from_user(&blk, cpu->lg->lguest_data->blocked_interrupts, |
d7e28ffe | 182 | sizeof(blk))) |
abd41f03 | 183 | return LGUEST_IRQS; |
177e449d | 184 | bitmap_andnot(blk, cpu->irqs_pending, blk, LGUEST_IRQS); |
d7e28ffe | 185 | |
bff672e6 | 186 | /* Find the first interrupt. */ |
d7e28ffe | 187 | irq = find_first_bit(blk, LGUEST_IRQS); |
a32a8813 | 188 | *more = find_next_bit(blk, LGUEST_IRQS, irq+1); |
abd41f03 RR |
189 | |
190 | return irq; | |
191 | } | |
192 | ||
2e04ef76 RR |
193 | /* |
194 | * This actually diverts the Guest to running an interrupt handler, once an | |
195 | * interrupt has been identified by interrupt_pending(). | |
196 | */ | |
a32a8813 | 197 | void try_deliver_interrupt(struct lg_cpu *cpu, unsigned int irq, bool more) |
abd41f03 RR |
198 | { |
199 | struct desc_struct *idt; | |
200 | ||
201 | BUG_ON(irq >= LGUEST_IRQS); | |
d7e28ffe | 202 | |
2e04ef76 RR |
203 | /* |
204 | * They may be in the middle of an iret, where they asked us never to | |
205 | * deliver interrupts. | |
206 | */ | |
382ac6b3 GOC |
207 | if (cpu->regs->eip >= cpu->lg->noirq_start && |
208 | (cpu->regs->eip < cpu->lg->noirq_end)) | |
d7e28ffe RR |
209 | return; |
210 | ||
bff672e6 | 211 | /* If they're halted, interrupts restart them. */ |
66686c2a | 212 | if (cpu->halted) { |
d7e28ffe | 213 | /* Re-enable interrupts. */ |
382ac6b3 GOC |
214 | if (put_user(X86_EFLAGS_IF, &cpu->lg->lguest_data->irq_enabled)) |
215 | kill_guest(cpu, "Re-enabling interrupts"); | |
66686c2a | 216 | cpu->halted = 0; |
d7e28ffe | 217 | } else { |
bff672e6 | 218 | /* Otherwise we check if they have interrupts disabled. */ |
d7e28ffe | 219 | u32 irq_enabled; |
382ac6b3 | 220 | if (get_user(irq_enabled, &cpu->lg->lguest_data->irq_enabled)) |
d7e28ffe | 221 | irq_enabled = 0; |
a32a8813 RR |
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); | |
d7e28ffe | 226 | return; |
a32a8813 | 227 | } |
d7e28ffe RR |
228 | } |
229 | ||
2e04ef76 RR |
230 | /* |
231 | * Look at the IDT entry the Guest gave us for this interrupt. The | |
bff672e6 | 232 | * first 32 (FIRST_EXTERNAL_VECTOR) entries are for traps, so we skip |
2e04ef76 RR |
233 | * over them. |
234 | */ | |
fc708b3e | 235 | idt = &cpu->arch.idt[FIRST_EXTERNAL_VECTOR+irq]; |
bff672e6 | 236 | /* If they don't have a handler (yet?), we just ignore it */ |
d7e28ffe | 237 | if (idt_present(idt->a, idt->b)) { |
bff672e6 | 238 | /* OK, mark it no longer pending and deliver it. */ |
177e449d | 239 | clear_bit(irq, cpu->irqs_pending); |
2e04ef76 RR |
240 | /* |
241 | * set_guest_interrupt() takes the interrupt descriptor and a | |
bff672e6 | 242 | * flag to say whether this interrupt pushes an error code onto |
2e04ef76 RR |
243 | * the stack as well: virtual interrupts never do. |
244 | */ | |
df1693ab | 245 | set_guest_interrupt(cpu, idt->a, idt->b, false); |
d7e28ffe | 246 | } |
6c8dca5d | 247 | |
2e04ef76 RR |
248 | /* |
249 | * Every time we deliver an interrupt, we update the timestamp in the | |
6c8dca5d RR |
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 | |
2e04ef76 RR |
253 | * timer interrupt. |
254 | */ | |
382ac6b3 | 255 | write_timestamp(cpu); |
a32a8813 | 256 | |
2e04ef76 RR |
257 | /* |
258 | * If there are no other interrupts we want to deliver, clear | |
259 | * the pending flag. | |
260 | */ | |
a32a8813 RR |
261 | if (!more) |
262 | put_user(0, &cpu->lg->lguest_data->irq_pending); | |
d7e28ffe | 263 | } |
9f155a9b RR |
264 | |
265 | /* And this is the routine when we want to set an interrupt for the Guest. */ | |
266 | void set_interrupt(struct lg_cpu *cpu, unsigned int irq) | |
267 | { | |
2e04ef76 RR |
268 | /* |
269 | * Next time the Guest runs, the core code will see if it can deliver | |
270 | * this interrupt. | |
271 | */ | |
9f155a9b RR |
272 | set_bit(irq, cpu->irqs_pending); |
273 | ||
2e04ef76 RR |
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 | */ | |
9f155a9b RR |
278 | if (!wake_up_process(cpu->tsk)) |
279 | kick_process(cpu->tsk); | |
280 | } | |
c18acd73 RR |
281 | /*:*/ |
282 | ||
2e04ef76 RR |
283 | /* |
284 | * Linux uses trap 128 for system calls. Plan9 uses 64, and Ron Minnich sent | |
c18acd73 RR |
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 | |
2e04ef76 RR |
289 | * userbase. Oh well. |
290 | */ | |
c18acd73 RR |
291 | static 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. */ | |
298 | bool 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 | ||
308 | int init_interrupts(void) | |
309 | { | |
310 | /* If they want some strange system call vector, reserve it now */ | |
b77b881f YL |
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); | |
c18acd73 | 319 | } |
b77b881f | 320 | |
c18acd73 RR |
321 | return 0; |
322 | } | |
323 | ||
324 | void free_interrupts(void) | |
325 | { | |
326 | if (syscall_vector != SYSCALL_VECTOR) | |
327 | clear_bit(syscall_vector, used_vectors); | |
328 | } | |
d7e28ffe | 329 | |
2e04ef76 RR |
330 | /*H:220 |
331 | * Now we've got the routines to deliver interrupts, delivering traps like | |
a6bd8e13 | 332 | * page fault is easy. The only trick is that Intel decided that some traps |
2e04ef76 RR |
333 | * should have error codes: |
334 | */ | |
df1693ab | 335 | static bool has_err(unsigned int trap) |
d7e28ffe RR |
336 | { |
337 | return (trap == 8 || (trap >= 10 && trap <= 14) || trap == 17); | |
338 | } | |
339 | ||
bff672e6 | 340 | /* deliver_trap() returns true if it could deliver the trap. */ |
df1693ab | 341 | bool deliver_trap(struct lg_cpu *cpu, unsigned int num) |
d7e28ffe | 342 | { |
2e04ef76 RR |
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 | */ | |
fc708b3e | 347 | if (num >= ARRAY_SIZE(cpu->arch.idt)) |
df1693ab | 348 | return false; |
d7e28ffe | 349 | |
2e04ef76 RR |
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 | */ | |
fc708b3e | 354 | if (!idt_present(cpu->arch.idt[num].a, cpu->arch.idt[num].b)) |
df1693ab | 355 | return false; |
fc708b3e GOC |
356 | set_guest_interrupt(cpu, cpu->arch.idt[num].a, |
357 | cpu->arch.idt[num].b, has_err(num)); | |
df1693ab | 358 | return true; |
d7e28ffe RR |
359 | } |
360 | ||
2e04ef76 RR |
361 | /*H:250 |
362 | * Here's the hard part: returning to the Host every time a trap happens | |
bff672e6 | 363 | * and then calling deliver_trap() and re-entering the Guest is slow. |
e1e72965 RR |
364 | * Particularly because Guest userspace system calls are traps (usually trap |
365 | * 128). | |
bff672e6 RR |
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 | |
e1e72965 | 371 | * the other hypervisors would beat it up at lunchtime. |
bff672e6 | 372 | * |
56adbe9d | 373 | * This routine indicates if a particular trap number could be delivered |
2e04ef76 RR |
374 | * directly. |
375 | */ | |
df1693ab | 376 | static bool direct_trap(unsigned int num) |
d7e28ffe | 377 | { |
2e04ef76 RR |
378 | /* |
379 | * Hardware interrupts don't go to the Guest at all (except system | |
380 | * call). | |
381 | */ | |
c18acd73 | 382 | if (num >= FIRST_EXTERNAL_VECTOR && !could_be_syscall(num)) |
df1693ab | 383 | return false; |
d7e28ffe | 384 | |
2e04ef76 RR |
385 | /* |
386 | * The Host needs to see page faults (for shadow paging and to save the | |
bff672e6 | 387 | * fault address), general protection faults (in/out emulation) and |
6d7a5d1e | 388 | * device not available (TS handling) and of course, the hypercall trap. |
2e04ef76 | 389 | */ |
6d7a5d1e | 390 | return num != 14 && num != 13 && num != 7 && num != LGUEST_TRAP_ENTRY; |
d7e28ffe | 391 | } |
f56a384e RR |
392 | /*:*/ |
393 | ||
2e04ef76 RR |
394 | /*M:005 |
395 | * The Guest has the ability to turn its interrupt gates into trap gates, | |
f56a384e RR |
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 | |
2e04ef76 RR |
399 | * within noirq_start and noirq_end, where it can safely disable interrupts. |
400 | */ | |
f56a384e | 401 | |
2e04ef76 RR |
402 | /*M:006 |
403 | * The Guests do not use the sysenter (fast system call) instruction, | |
f56a384e RR |
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 | |
2e04ef76 RR |
408 | * entry.S |
409 | :*/ | |
d7e28ffe | 410 | |
2e04ef76 RR |
411 | /*H:260 |
412 | * When we make traps go directly into the Guest, we need to make sure | |
bff672e6 RR |
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 | * | |
2e04ef76 RR |
418 | * Which is deeply unfair, because (literally!) it wasn't the Guests' fault. |
419 | */ | |
4665ac8e | 420 | void pin_stack_pages(struct lg_cpu *cpu) |
d7e28ffe RR |
421 | { |
422 | unsigned int i; | |
423 | ||
2e04ef76 RR |
424 | /* |
425 | * Depending on the CONFIG_4KSTACKS option, the Guest can have one or | |
426 | * two pages of stack space. | |
427 | */ | |
382ac6b3 | 428 | for (i = 0; i < cpu->lg->stack_pages; i++) |
2e04ef76 RR |
429 | /* |
430 | * The stack grows *upwards*, so the address we're given is the | |
8057d763 RR |
431 | * start of the page after the kernel stack. Subtract one to |
432 | * get back onto the first stack page, and keep subtracting to | |
2e04ef76 RR |
433 | * get to the rest of the stack pages. |
434 | */ | |
1713608f | 435 | pin_page(cpu, cpu->esp1 - 1 - i * PAGE_SIZE); |
d7e28ffe RR |
436 | } |
437 | ||
2e04ef76 RR |
438 | /* |
439 | * Direct traps also mean that we need to know whenever the Guest wants to use | |
9f54288d RR |
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 | |
bff672e6 RR |
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 | |
2e04ef76 RR |
446 | * change stacks on each context switch. |
447 | */ | |
4665ac8e | 448 | void guest_set_stack(struct lg_cpu *cpu, u32 seg, u32 esp, unsigned int pages) |
d7e28ffe | 449 | { |
2e04ef76 RR |
450 | /* |
451 | * You're not allowed a stack segment with privilege level 0: bad Guest! | |
452 | */ | |
d7e28ffe | 453 | if ((seg & 0x3) != GUEST_PL) |
382ac6b3 | 454 | kill_guest(cpu, "bad stack segment %i", seg); |
bff672e6 | 455 | /* We only expect one or two stack pages. */ |
d7e28ffe | 456 | if (pages > 2) |
382ac6b3 | 457 | kill_guest(cpu, "bad stack pages %u", pages); |
bff672e6 | 458 | /* Save where the stack is, and how many pages */ |
4665ac8e GOC |
459 | cpu->ss1 = seg; |
460 | cpu->esp1 = esp; | |
461 | cpu->lg->stack_pages = pages; | |
bff672e6 | 462 | /* Make sure the new stack pages are mapped */ |
4665ac8e | 463 | pin_stack_pages(cpu); |
d7e28ffe RR |
464 | } |
465 | ||
2e04ef76 RR |
466 | /* |
467 | * All this reference to mapping stacks leads us neatly into the other complex | |
468 | * part of the Host: page table handling. | |
469 | */ | |
bff672e6 | 470 | |
2e04ef76 RR |
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 | */ | |
382ac6b3 | 475 | static void set_trap(struct lg_cpu *cpu, struct desc_struct *trap, |
d7e28ffe RR |
476 | unsigned int num, u32 lo, u32 hi) |
477 | { | |
478 | u8 type = idt_type(lo, hi); | |
479 | ||
bff672e6 | 480 | /* We zero-out a not-present entry */ |
d7e28ffe RR |
481 | if (!idt_present(lo, hi)) { |
482 | trap->a = trap->b = 0; | |
483 | return; | |
484 | } | |
485 | ||
bff672e6 | 486 | /* We only support interrupt and trap gates. */ |
d7e28ffe | 487 | if (type != 0xE && type != 0xF) |
382ac6b3 | 488 | kill_guest(cpu, "bad IDT type %i", type); |
d7e28ffe | 489 | |
2e04ef76 RR |
490 | /* |
491 | * We only copy the handler address, present bit, privilege level and | |
bff672e6 RR |
492 | * type. The privilege level controls where the trap can be triggered |
493 | * manually with an "int" instruction. This is usually GUEST_PL, | |
2e04ef76 RR |
494 | * except for system calls which userspace can use. |
495 | */ | |
d7e28ffe RR |
496 | trap->a = ((__KERNEL_CS|GUEST_PL)<<16) | (lo&0x0000FFFF); |
497 | trap->b = (hi&0xFFFFEF00); | |
498 | } | |
499 | ||
2e04ef76 RR |
500 | /*H:230 |
501 | * While we're here, dealing with delivering traps and interrupts to the | |
bff672e6 RR |
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 | |
2e04ef76 RR |
507 | * LHCALL_LOAD_IDT_ENTRY hypercall before: that comes here. |
508 | */ | |
fc708b3e | 509 | void load_guest_idt_entry(struct lg_cpu *cpu, unsigned int num, u32 lo, u32 hi) |
d7e28ffe | 510 | { |
2e04ef76 RR |
511 | /* |
512 | * Guest never handles: NMI, doublefault, spurious interrupt or | |
513 | * hypercall. We ignore when it tries to set them. | |
514 | */ | |
d7e28ffe RR |
515 | if (num == 2 || num == 8 || num == 15 || num == LGUEST_TRAP_ENTRY) |
516 | return; | |
517 | ||
2e04ef76 RR |
518 | /* |
519 | * Mark the IDT as changed: next time the Guest runs we'll know we have | |
520 | * to copy this again. | |
521 | */ | |
ae3749dc | 522 | cpu->changed |= CHANGED_IDT; |
bff672e6 | 523 | |
56adbe9d | 524 | /* Check that the Guest doesn't try to step outside the bounds. */ |
fc708b3e | 525 | if (num >= ARRAY_SIZE(cpu->arch.idt)) |
382ac6b3 | 526 | kill_guest(cpu, "Setting idt entry %u", num); |
56adbe9d | 527 | else |
382ac6b3 | 528 | set_trap(cpu, &cpu->arch.idt[num], num, lo, hi); |
d7e28ffe RR |
529 | } |
530 | ||
2e04ef76 RR |
531 | /* |
532 | * The default entry for each interrupt points into the Switcher routines which | |
bff672e6 | 533 | * simply return to the Host. The run_guest() loop will then call |
2e04ef76 RR |
534 | * deliver_trap() to bounce it back into the Guest. |
535 | */ | |
d7e28ffe RR |
536 | static void default_idt_entry(struct desc_struct *idt, |
537 | int trap, | |
0c12091d RR |
538 | const unsigned long handler, |
539 | const struct desc_struct *base) | |
d7e28ffe | 540 | { |
bff672e6 | 541 | /* A present interrupt gate. */ |
d7e28ffe RR |
542 | u32 flags = 0x8e00; |
543 | ||
2e04ef76 RR |
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 | */ | |
d7e28ffe RR |
548 | if (trap == LGUEST_TRAP_ENTRY) |
549 | flags |= (GUEST_PL << 13); | |
0c12091d | 550 | else if (base) |
2e04ef76 RR |
551 | /* |
552 | * Copy privilege level from what Guest asked for. This allows | |
553 | * debug (int 3) traps from Guest userspace, for example. | |
554 | */ | |
0c12091d | 555 | flags |= (base->b & 0x6000); |
d7e28ffe | 556 | |
bff672e6 | 557 | /* Now pack it into the IDT entry in its weird format. */ |
d7e28ffe RR |
558 | idt->a = (LGUEST_CS<<16) | (handler&0x0000FFFF); |
559 | idt->b = (handler&0xFFFF0000) | flags; | |
560 | } | |
561 | ||
bff672e6 | 562 | /* When the Guest first starts, we put default entries into the IDT. */ |
d7e28ffe RR |
563 | void 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++) | |
0c12091d | 569 | default_idt_entry(&state->guest_idt[i], i, def[i], NULL); |
d7e28ffe RR |
570 | } |
571 | ||
2e04ef76 RR |
572 | /*H:240 |
573 | * We don't use the IDT entries in the "struct lguest" directly, instead | |
bff672e6 | 574 | * we copy them into the IDT which we've set up for Guests on this CPU, just |
2e04ef76 RR |
575 | * before we run the Guest. This routine does that copy. |
576 | */ | |
fc708b3e | 577 | void copy_traps(const struct lg_cpu *cpu, struct desc_struct *idt, |
d7e28ffe RR |
578 | const unsigned long *def) |
579 | { | |
580 | unsigned int i; | |
581 | ||
2e04ef76 RR |
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 | */ | |
fc708b3e | 586 | for (i = 0; i < ARRAY_SIZE(cpu->arch.idt); i++) { |
0c12091d RR |
587 | const struct desc_struct *gidt = &cpu->arch.idt[i]; |
588 | ||
56adbe9d RR |
589 | /* If no Guest can ever override this trap, leave it alone. */ |
590 | if (!direct_trap(i)) | |
591 | continue; | |
592 | ||
2e04ef76 RR |
593 | /* |
594 | * Only trap gates (type 15) can go direct to the Guest. | |
56adbe9d RR |
595 | * Interrupt gates (type 14) disable interrupts as they are |
596 | * entered, which we never let the Guest do. Not present | |
0c12091d RR |
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 | |
2e04ef76 RR |
601 | * interrupt. |
602 | */ | |
0c12091d RR |
603 | if (idt_type(gidt->a, gidt->b) == 0xF) |
604 | idt[i] = *gidt; | |
d7e28ffe | 605 | else |
0c12091d | 606 | default_idt_entry(&idt[i], i, def[i], gidt); |
d7e28ffe | 607 | } |
d7e28ffe RR |
608 | } |
609 | ||
e1e72965 RR |
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 | * | |
2e04ef76 RR |
621 | * 0 means "turn off the clock". |
622 | */ | |
ad8d8f3b | 623 | void guest_set_clockevent(struct lg_cpu *cpu, unsigned long delta) |
d7e28ffe RR |
624 | { |
625 | ktime_t expires; | |
626 | ||
627 | if (unlikely(delta == 0)) { | |
628 | /* Clock event device is shutting down. */ | |
ad8d8f3b | 629 | hrtimer_cancel(&cpu->hrt); |
d7e28ffe RR |
630 | return; |
631 | } | |
632 | ||
2e04ef76 RR |
633 | /* |
634 | * We use wallclock time here, so the Guest might not be running for | |
e1e72965 | 635 | * all the time between now and the timer interrupt it asked for. This |
2e04ef76 RR |
636 | * is almost always the right thing to do. |
637 | */ | |
d7e28ffe | 638 | expires = ktime_add_ns(ktime_get_real(), delta); |
ad8d8f3b | 639 | hrtimer_start(&cpu->hrt, expires, HRTIMER_MODE_ABS); |
d7e28ffe RR |
640 | } |
641 | ||
e1e72965 | 642 | /* This is the function called when the Guest's timer expires. */ |
d7e28ffe RR |
643 | static enum hrtimer_restart clockdev_fn(struct hrtimer *timer) |
644 | { | |
ad8d8f3b | 645 | struct lg_cpu *cpu = container_of(timer, struct lg_cpu, hrt); |
d7e28ffe | 646 | |
e1e72965 | 647 | /* Remember the first interrupt is the timer interrupt. */ |
9f155a9b | 648 | set_interrupt(cpu, 0); |
d7e28ffe RR |
649 | return HRTIMER_NORESTART; |
650 | } | |
651 | ||
e1e72965 | 652 | /* This sets up the timer for this Guest. */ |
ad8d8f3b | 653 | void init_clockdev(struct lg_cpu *cpu) |
d7e28ffe | 654 | { |
ad8d8f3b GOC |
655 | hrtimer_init(&cpu->hrt, CLOCK_REALTIME, HRTIMER_MODE_ABS); |
656 | cpu->hrt.function = clockdev_fn; | |
d7e28ffe | 657 | } |